feat: Add Face Parsing model BiSeNet model trained on CelebMask dataset (#35)

* Add BiSeNet face parsing implementation

* Add parsing model weights configuration

* Export BiSeNet in main package

* Add face parsing tests

* Add face parsing examples and script

* Bump version to 1.5.0

* Update documentation for face parsing

* Fix face parsing notebook to use lips instead of mouth

* chore: Update the face parsing example

* fix: Fix model argument to use Enum

* ref: Move vis_parsing_map function into visualization.py

* docs: Update README.md

.github/workflows/build.yml

@@ -1,64 +0,0 @@
-name: Build, Test, and Publish
-
-on:
-  push:
-    branches:
-      - main
-    tags:
-      - "v*.*.*"  # Trigger publish on version tags
-  pull_request:
-    branches:
-      - main
-
-jobs:
-  build:
-    runs-on: ubuntu-latest
-
-    strategy:
-      matrix:
-        python-version: ["3.8", "3.9", "3.10"]
-
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v3
-
-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: ${{ matrix.python-version }}
-
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          python -m pip install .[dev] || pip install pytest  # Use extras_require if available
-
-      - name: Run Tests
-        run: |
-          pytest
-
-  publish:
-    runs-on: ubuntu-latest
-    needs: build  # Publish only if tests pass
-
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v3
-
-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: "3.10"  # Use a single Python version for publishing
-
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          python -m pip install build twine
-
-      - name: Build Package
-        run: python -m build
-
-      - name: Publish to PyPI
-        env:
-          TWINE_USERNAME: __token__
-          TWINE_PASSWORD: ${{ secrets.PYPI_API_TOKEN }}
-        run: twine upload dist/*

.github/workflows/ci.yml

@@ -0,0 +1,87 @@
+name: CI
+
+on:
+  push:
+    branches:
+      - main
+      - develop
+  pull_request:
+    branches:
+      - main
+      - develop
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.10", "3.11", "3.12", "3.13"]
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set up Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install .[dev]
+
+      - name: Check ONNX Runtime providers
+        run: |
+          python -c "import onnxruntime as ort; print('Available providers:', ort.get_available_providers())"
+
+      - name: Lint with ruff (if available)
+        run: |
+          pip install ruff || true
+          ruff check . --exit-zero || true
+        continue-on-error: true
+
+      - name: Run tests
+        run: pytest -v --tb=short
+
+      - name: Test package imports
+        run: |
+          python -c "from uniface import RetinaFace, ArcFace, Landmark106, AgeGender; print('All imports successful')"
+
+  build:
+    runs-on: ubuntu-latest
+    needs: test
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+          cache: 'pip'
+
+      - name: Install build tools
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install build
+
+      - name: Build package
+        run: python -m build
+
+      - name: Check package
+        run: |
+          python -m pip install twine
+          twine check dist/*
+
+      - name: Upload build artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          name: dist-python-${{ github.sha }}
+          path: dist/
+          retention-days: 7
+

.github/workflows/publish.yml

@@ -0,0 +1,108 @@
+name: Publish to PyPI
+
+on:
+  push:
+    tags:
+      - "v*.*.*"  # Trigger only on version tags like v0.1.9
+
+jobs:
+  validate:
+    runs-on: ubuntu-latest
+    outputs:
+      version: ${{ steps.get_version.outputs.version }}
+      tag_version: ${{ steps.get_version.outputs.tag_version }}
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Get version from tag and pyproject.toml
+        id: get_version
+        run: |
+          TAG_VERSION=${GITHUB_REF#refs/tags/v}
+          echo "tag_version=$TAG_VERSION" >> $GITHUB_OUTPUT
+
+          PYPROJECT_VERSION=$(grep -Po '(?<=^version = ")[^"]*' pyproject.toml)
+          echo "version=$PYPROJECT_VERSION" >> $GITHUB_OUTPUT
+
+          echo "Tag version: v$TAG_VERSION"
+          echo "pyproject.toml version: $PYPROJECT_VERSION"
+
+      - name: Verify version match
+        run: |
+          if [ "${{ steps.get_version.outputs.tag_version }}" != "${{ steps.get_version.outputs.version }}" ]; then
+            echo "Error: Tag version (${{ steps.get_version.outputs.tag_version }}) does not match pyproject.toml version (${{ steps.get_version.outputs.version }})"
+            exit 1
+          fi
+          echo "Version validation passed: ${{ steps.get_version.outputs.version }}"
+
+  test:
+    runs-on: ubuntu-latest
+    needs: validate
+
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.10", "3.11", "3.12", "3.13"]
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set up Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install .[dev]
+
+      - name: Run tests
+        run: pytest -v
+
+  publish:
+    runs-on: ubuntu-latest
+    needs: [validate, test]
+    permissions:
+      contents: write
+      id-token: write
+    environment:
+      name: pypi
+      url: https://pypi.org/project/uniface/
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+          cache: 'pip'
+
+      - name: Install build tools
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install build twine
+
+      - name: Build package
+        run: python -m build
+
+      - name: Check package
+        run: twine check dist/*
+
+      - name: Publish to PyPI
+        env:
+          TWINE_USERNAME: __token__
+          TWINE_PASSWORD: ${{ secrets.PYPI_API_TOKEN }}
+        run: twine upload dist/*
+
+      - name: Create GitHub Release
+        uses: softprops/action-gh-release@v1
+        with:
+          files: dist/*
+          generate_release_notes: true
+

.gitignore

@@ -1,3 +1,5 @@
+tmp_*
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]

CONTRIBUTING.md

@@ -0,0 +1,63 @@
+# Contributing to UniFace
+
+Thank you for considering contributing to UniFace! We welcome contributions of all kinds.
+
+## How to Contribute
+
+### Reporting Issues
+
+- Use GitHub Issues to report bugs or suggest features
+- Include clear descriptions and reproducible examples
+- Check existing issues before creating new ones
+
+### Pull Requests
+
+1. Fork the repository
+2. Create a new branch for your feature
+3. Write clear, documented code with type hints
+4. Add tests for new functionality
+5. Ensure all tests pass
+6. Submit a pull request with a clear description
+
+### Code Style
+
+- Follow PEP8 guidelines
+- Use type hints (Python 3.10+)
+- Write docstrings for public APIs
+- Keep code simple and readable
+
+## Development Setup
+
+```bash
+git clone https://github.com/yakhyo/uniface.git
+cd uniface
+pip install -e ".[dev]"
+```
+
+## Running Tests
+
+```bash
+pytest tests/
+```
+
+## Examples
+
+Example notebooks demonstrating library usage:
+
+| Example | Notebook |
+|---------|----------|
+| Face Detection | [face_detection.ipynb](examples/face_detection.ipynb) |
+| Face Alignment | [face_alignment.ipynb](examples/face_alignment.ipynb) |
+| Face Recognition | [face_analyzer.ipynb](examples/face_analyzer.ipynb) |
+| Face Verification | [face_verification.ipynb](examples/face_verification.ipynb) |
+| Face Search | [face_search.ipynb](examples/face_search.ipynb) |
+
+## Questions?
+
+Open an issue or start a discussion on GitHub.
+
+
+
+
+
+

MODELS.md

@@ -0,0 +1,457 @@
+# UniFace Model Zoo
+
+Complete guide to all available models, their performance characteristics, and selection criteria.
+
+---
+
+## Face Detection Models
+
+### RetinaFace Family
+
+RetinaFace models are trained on the WIDER FACE dataset and provide excellent accuracy-speed tradeoffs.
+
+| Model Name     | Params | Size  | Easy   | Medium | Hard   | Use Case                      |
+| -------------- | ------ | ----- | ------ | ------ | ------ | ----------------------------- |
+| `MNET_025`   | 0.4M   | 1.7MB | 88.48% | 87.02% | 80.61% | Mobile/Edge devices           |
+| `MNET_050`   | 1.0M   | 2.6MB | 89.42% | 87.97% | 82.40% | Mobile/Edge devices           |
+| `MNET_V1`    | 3.5M   | 3.8MB | 90.59% | 89.14% | 84.13% | Balanced mobile               |
+| `MNET_V2` ⭐ | 3.2M   | 3.5MB | 91.70% | 91.03% | 86.60% | **Recommended default** |
+| `RESNET18`   | 11.7M  | 27MB  | 92.50% | 91.02% | 86.63% | Server/High accuracy          |
+| `RESNET34`   | 24.8M  | 56MB  | 94.16% | 93.12% | 88.90% | Maximum accuracy              |
+
+**Accuracy**: WIDER FACE validation set (Easy/Medium/Hard subsets) - from [RetinaFace paper](https://arxiv.org/abs/1905.00641)
+**Speed**: Benchmark on your own hardware using `scripts/run_detection.py --iterations 100`
+
+#### Usage
+
+```python
+from uniface import RetinaFace
+from uniface.constants import RetinaFaceWeights
+
+# Default (recommended)
+detector = RetinaFace()  # Uses MNET_V2
+
+# Specific model
+detector = RetinaFace(
+    model_name=RetinaFaceWeights.MNET_025,  # Fastest
+    conf_thresh=0.5,
+    nms_thresh=0.4,
+    input_size=(640, 640)
+)
+```
+
+---
+
+### SCRFD Family
+
+SCRFD (Sample and Computation Redistribution for Efficient Face Detection) models offer state-of-the-art speed-accuracy tradeoffs.
+
+| Model Name       | Params | Size  | Easy   | Medium | Hard   | Use Case                        |
+| ---------------- | ------ | ----- | ------ | ------ | ------ | ------------------------------- |
+| `SCRFD_500M`   | 0.6M   | 2.5MB | 90.57% | 88.12% | 68.51% | Real-time applications          |
+| `SCRFD_10G` ⭐ | 4.2M   | 17MB  | 95.16% | 93.87% | 83.05% | **High accuracy + speed** |
+
+**Accuracy**: WIDER FACE validation set - from [SCRFD paper](https://arxiv.org/abs/2105.04714)
+**Speed**: Benchmark on your own hardware using `scripts/run_detection.py --iterations 100`
+
+#### Usage
+
+```python
+from uniface import SCRFD
+from uniface.constants import SCRFDWeights
+
+# Fast real-time detection
+detector = SCRFD(
+    model_name=SCRFDWeights.SCRFD_500M_KPS,
+    conf_thresh=0.5,
+    input_size=(640, 640)
+)
+
+# High accuracy
+detector = SCRFD(
+    model_name=SCRFDWeights.SCRFD_10G_KPS,
+    conf_thresh=0.5
+)
+```
+
+---
+
+### YOLOv5-Face Family
+
+YOLOv5-Face models provide excellent detection accuracy with 5-point facial landmarks, optimized for real-time applications.
+
+| Model Name     | Size | Easy   | Medium | Hard   | Use Case                       |
+| -------------- | ---- | ------ | ------ | ------ | ------------------------------ |
+| `YOLOV5N`    | 11MB | 93.61% | 91.52% | 80.53% | Lightweight/Mobile             |
+| `YOLOV5S` ⭐ | 28MB | 94.33% | 92.61% | 83.15% | **Real-time + accuracy** |
+| `YOLOV5M`    | 82MB | 95.30% | 93.76% | 85.28% | High accuracy                  |
+
+**Accuracy**: WIDER FACE validation set - from [YOLOv5-Face paper](https://arxiv.org/abs/2105.12931)
+**Speed**: Benchmark on your own hardware using `scripts/run_detection.py --iterations 100`
+**Note**: Fixed input size of 640×640. Models exported to ONNX from [deepcam-cn/yolov5-face](https://github.com/deepcam-cn/yolov5-face)
+
+#### Usage
+
+```python
+from uniface import YOLOv5Face
+from uniface.constants import YOLOv5FaceWeights
+
+# Lightweight/Mobile
+detector = YOLOv5Face(
+    model_name=YOLOv5FaceWeights.YOLOV5N,
+    conf_thresh=0.6,
+    nms_thresh=0.5
+)
+
+# Real-time detection (recommended)
+detector = YOLOv5Face(
+    model_name=YOLOv5FaceWeights.YOLOV5S,
+    conf_thresh=0.6,
+    nms_thresh=0.5
+)
+
+# High accuracy
+detector = YOLOv5Face(
+    model_name=YOLOv5FaceWeights.YOLOV5M,
+    conf_thresh=0.6
+)
+
+# Detect faces with landmarks
+faces = detector.detect(image)
+for face in faces:
+    bbox = face['bbox']  # [x1, y1, x2, y2]
+    confidence = face['confidence']
+    landmarks = face['landmarks']  # 5-point landmarks (5, 2)
+```
+
+---
+
+## Face Recognition Models
+
+### ArcFace
+
+State-of-the-art face recognition using additive angular margin loss.
+
+| Model Name  | Backbone  | Params | Size  | Use Case                         |
+| ----------- | --------- | ------ | ----- | -------------------------------- |
+| `MNET` ⭐ | MobileNet | 2.0M   | 8MB   | **Balanced (recommended)** |
+| `RESNET`  | ResNet50  | 43.6M  | 166MB | Maximum accuracy                 |
+
+**Dataset**: Trained on MS1M-V2 (5.8M images, 85K identities)
+**Accuracy**: Benchmark on your own dataset or use standard face verification benchmarks
+
+#### Usage
+
+```python
+from uniface import ArcFace
+from uniface.constants import ArcFaceWeights
+
+# Default (MobileNet backbone)
+recognizer = ArcFace()
+
+# High accuracy (ResNet50 backbone)
+recognizer = ArcFace(model_name=ArcFaceWeights.RESNET)
+
+# Extract embedding
+embedding = recognizer.get_normalized_embedding(image, landmarks)
+# Returns: (1, 512) normalized embedding vector
+```
+
+---
+
+### MobileFace
+
+Lightweight face recognition optimized for mobile devices.
+
+| Model Name        | Backbone         | Params | Size | LFW    | CALFW  | CPLFW  | AgeDB-30 | Use Case              |
+| ----------------- | ---------------- | ------ | ---- | ------ | ------ | ------ | -------- | --------------------- |
+| `MNET_025`      | MobileNetV1 0.25 | 0.36M  | 1MB  | 98.76% | 92.02% | 82.37% | 90.02%   | Ultra-lightweight     |
+| `MNET_V2` ⭐    | MobileNetV2      | 2.29M  | 4MB  | 99.55% | 94.87% | 86.89% | 95.16%   | **Mobile/Edge** |
+| `MNET_V3_SMALL` | MobileNetV3-S    | 1.25M  | 3MB  | 99.30% | 93.77% | 85.29% | 92.79%   | Mobile optimized      |
+| `MNET_V3_LARGE` | MobileNetV3-L    | 3.52M  | 10MB | 99.53% | 94.56% | 86.79% | 95.13%   | Balanced mobile       |
+
+**Dataset**: Trained on MS1M-V2 (5.8M images, 85K identities)
+**Accuracy**: Evaluated on LFW, CALFW, CPLFW, and AgeDB-30 benchmarks
+**Note**: These models are lightweight alternatives to ArcFace for resource-constrained environments
+
+#### Usage
+
+```python
+from uniface import MobileFace
+from uniface.constants import MobileFaceWeights
+
+# Lightweight
+recognizer = MobileFace(model_name=MobileFaceWeights.MNET_V2)
+```
+
+---
+
+### SphereFace
+
+Face recognition using angular softmax loss.
+
+| Model Name   | Backbone | Params | Size | LFW    | CALFW  | CPLFW  | AgeDB-30 | Use Case            |
+| ------------ | -------- | ------ | ---- | ------ | ------ | ------ | -------- | ------------------- |
+| `SPHERE20` | Sphere20 | 24.5M  | 50MB | 99.67% | 95.61% | 88.75% | 96.58%   | Research/Comparison |
+| `SPHERE36` | Sphere36 | 34.6M  | 92MB | 99.72% | 95.64% | 89.92% | 96.83%   | Research/Comparison |
+
+**Dataset**: Trained on MS1M-V2 (5.8M images, 85K identities)
+**Accuracy**: Evaluated on LFW, CALFW, CPLFW, and AgeDB-30 benchmarks
+**Note**: SphereFace uses angular softmax loss, an earlier approach before ArcFace. These models provide good accuracy with moderate resource requirements.
+
+#### Usage
+
+```python
+from uniface import SphereFace
+from uniface.constants import SphereFaceWeights
+
+recognizer = SphereFace(model_name=SphereFaceWeights.SPHERE20)
+```
+
+---
+
+## Facial Landmark Models
+
+### 106-Point Landmark Detection
+
+High-precision facial landmark localization.
+
+| Model Name | Points | Params | Size | Use Case                 |
+| ---------- | ------ | ------ | ---- | ------------------------ |
+| `2D106`  | 106    | 3.7M   | 14MB | Face alignment, analysis |
+
+**Note**: Provides 106 facial keypoints for detailed face analysis and alignment
+
+#### Usage
+
+```python
+from uniface import Landmark106
+
+landmarker = Landmark106()
+landmarks = landmarker.get_landmarks(image, bbox)
+# Returns: (106, 2) array of (x, y) coordinates
+```
+
+**Landmark Groups:**
+
+- Face contour: 0-32 (33 points)
+- Eyebrows: 33-50 (18 points)
+- Nose: 51-62 (12 points)
+- Eyes: 63-86 (24 points)
+- Mouth: 87-105 (19 points)
+
+---
+
+## Attribute Analysis Models
+
+### Age & Gender Detection
+
+| Model Name  | Attributes  | Params | Size | Use Case        |
+| ----------- | ----------- | ------ | ---- | --------------- |
+| `DEFAULT` | Age, Gender | 2.1M   | 8MB  | General purpose |
+
+**Dataset**: Trained on CelebA
+**Note**: Accuracy varies by demographic and image quality. Test on your specific use case.
+
+#### Usage
+
+```python
+from uniface import AgeGender
+
+predictor = AgeGender()
+gender, age = predictor.predict(image, bbox)
+# Returns: (gender, age_in_years)
+# gender: 0 for Female, 1 for Male
+```
+
+---
+
+### Emotion Detection
+
+| Model Name    | Classes | Params | Size | Use Case        |
+| ------------- | ------- | ------ | ---- | --------------- |
+| `AFFECNET7` | 7       | 0.5M   | 2MB  | 7-class emotion |
+| `AFFECNET8` | 8       | 0.5M   | 2MB  | 8-class emotion |
+
+**Classes (7)**: Neutral, Happy, Sad, Surprise, Fear, Disgust, Anger
+**Classes (8)**: Above + Contempt
+
+**Dataset**: Trained on AffectNet
+**Note**: Emotion detection accuracy depends heavily on facial expression clarity and cultural context
+
+#### Usage
+
+```python
+from uniface import Emotion
+from uniface.constants import DDAMFNWeights
+
+predictor = Emotion(model_name=DDAMFNWeights.AFFECNET7)
+emotion, confidence = predictor.predict(image, landmarks)
+```
+
+---
+
+## Gaze Estimation Models
+
+### MobileGaze Family
+
+Real-time gaze direction prediction models trained on Gaze360 dataset. Returns pitch (vertical) and yaw (horizontal) angles in radians.
+
+| Model Name     | Params | Size    | MAE*  | Use Case                      |
+| -------------- | ------ | ------- | ----- | ----------------------------- |
+| `RESNET18`   | 11.7M  | 43 MB   | 12.84 | Balanced accuracy/speed       |
+| `RESNET34` ⭐ | 24.8M  | 81.6 MB | 11.33 | **Recommended default** |
+| `RESNET50`   | 25.6M  | 91.3 MB | 11.34 | High accuracy                 |
+| `MOBILENET_V2` | 3.5M   | 9.59 MB | 13.07 | Mobile/Edge devices           |
+| `MOBILEONE_S0` | 2.1M   | 4.8 MB  | 12.58 | Lightweight/Real-time         |
+
+*MAE (Mean Absolute Error) in degrees on Gaze360 test set - lower is better
+
+**Dataset**: Trained on Gaze360 (indoor/outdoor scenes with diverse head poses)
+**Training**: 200 epochs with classification-based approach (binned angles)
+
+#### Usage
+
+```python
+from uniface import MobileGaze
+from uniface.constants import GazeWeights
+import numpy as np
+
+# Default (recommended)
+gaze_estimator = MobileGaze()  # Uses RESNET34
+
+# Lightweight model
+gaze_estimator = MobileGaze(model_name=GazeWeights.MOBILEONE_S0)
+
+# Estimate gaze from face crop
+pitch, yaw = gaze_estimator.estimate(face_crop)
+print(f"Pitch: {np.degrees(pitch):.1f}°, Yaw: {np.degrees(yaw):.1f}°")
+```
+
+**Note**: Requires face crop as input. Use face detection first to obtain bounding boxes.
+
+---
+
+## Face Parsing Models
+
+### BiSeNet Family
+
+BiSeNet (Bilateral Segmentation Network) models for semantic face parsing. Segments face images into 19 facial component classes.
+
+| Model Name     | Params | Size    | Classes | Use Case                      |
+| -------------- | ------ | ------- | ------- | ----------------------------- |
+| `RESNET18` ⭐ | 13.3M  | 50.7 MB | 19      | **Recommended default** |
+| `RESNET34`   | 24.1M  | 89.2 MB | 19      | Higher accuracy               |
+
+**19 Facial Component Classes:**
+1. Background
+2. Skin
+3. Left Eyebrow
+4. Right Eyebrow
+5. Left Eye
+6. Right Eye
+7. Eye Glasses
+8. Left Ear
+9. Right Ear
+10. Ear Ring
+11. Nose
+12. Mouth
+13. Upper Lip
+14. Lower Lip
+15. Neck
+16. Neck Lace
+17. Cloth
+18. Hair
+19. Hat
+
+**Dataset**: Trained on CelebAMask-HQ
+**Architecture**: BiSeNet with ResNet backbone
+**Input Size**: 512×512 (automatically resized)
+
+#### Usage
+
+```python
+from uniface.parsing import BiSeNet
+from uniface.constants import ParsingWeights
+from uniface.visualization import vis_parsing_maps
+import cv2
+
+# Default (recommended)
+parser = BiSeNet()  # Uses RESNET18
+
+# Higher accuracy model
+parser = BiSeNet(model_name=ParsingWeights.RESNET34)
+
+# Parse face image (already cropped)
+mask = parser.parse(face_image)
+
+# Visualize with overlay
+face_rgb = cv2.cvtColor(face_image, cv2.COLOR_BGR2RGB)
+vis_result = vis_parsing_maps(face_rgb, mask, save_image=False)
+
+# mask shape: (H, W) with values 0-18 representing classes
+print(f"Detected {len(np.unique(mask))} facial components")
+```
+
+**Applications:**
+- Face makeup and beauty applications
+- Virtual try-on systems
+- Face editing and manipulation
+- Facial feature extraction
+- Portrait segmentation
+
+**Note**: Input should be a cropped face image. For full pipeline, use face detection first to obtain face crops.
+
+---
+
+## Model Updates
+
+Models are automatically downloaded and cached on first use. Cache location: `~/.uniface/models/`
+
+### Manual Model Management
+
+```python
+from uniface.model_store import verify_model_weights
+from uniface.constants import RetinaFaceWeights
+
+# Download specific model
+model_path = verify_model_weights(
+    RetinaFaceWeights.MNET_V2,
+    root='./custom_cache'
+)
+
+# Models are verified with SHA-256 checksums
+```
+
+### Download All Models
+
+```bash
+# Using the provided script
+python scripts/download_model.py
+
+# Download specific model
+python scripts/download_model.py --model MNET_V2
+```
+
+---
+
+## References
+
+### Model Training & Architectures
+
+- **RetinaFace Training**: [yakhyo/retinaface-pytorch](https://github.com/yakhyo/retinaface-pytorch) - PyTorch implementation and training code
+- **YOLOv5-Face Original**: [deepcam-cn/yolov5-face](https://github.com/deepcam-cn/yolov5-face) - Original PyTorch implementation
+- **YOLOv5-Face ONNX**: [yakhyo/yolov5-face-onnx-inference](https://github.com/yakhyo/yolov5-face-onnx-inference) - ONNX inference implementation
+- **Face Recognition Training**: [yakhyo/face-recognition](https://github.com/yakhyo/face-recognition) - ArcFace, MobileFace, SphereFace training code
+- **Gaze Estimation Training**: [yakhyo/gaze-estimation](https://github.com/yakhyo/gaze-estimation) - MobileGaze training code and pretrained weights
+- **Face Parsing Training**: [yakhyo/face-parsing](https://github.com/yakhyo/face-parsing) - BiSeNet training code and pretrained weights
+- **InsightFace**: [deepinsight/insightface](https://github.com/deepinsight/insightface) - Model architectures and pretrained weights
+
+### Papers
+
+- **RetinaFace**: [Single-Shot Multi-Level Face Localisation in the Wild](https://arxiv.org/abs/1905.00641)
+- **SCRFD**: [Sample and Computation Redistribution for Efficient Face Detection](https://arxiv.org/abs/2105.04714)
+- **YOLOv5-Face**: [YOLO5Face: Why Reinventing a Face Detector](https://arxiv.org/abs/2105.12931)
+- **ArcFace**: [Additive Angular Margin Loss for Deep Face Recognition](https://arxiv.org/abs/1801.07698)
+- **SphereFace**: [Deep Hypersphere Embedding for Face Recognition](https://arxiv.org/abs/1704.08063)
+- **BiSeNet**: [Bilateral Segmentation Network for Real-time Semantic Segmentation](https://arxiv.org/abs/1808.00897)

QUICKSTART.md

@@ -0,0 +1,522 @@
+# UniFace Quick Start Guide
+
+Get up and running with UniFace in 5 minutes! This guide covers the most common use cases.
+
+---
+
+## Installation
+
+```bash
+# macOS (Apple Silicon) - automatically includes ARM64 optimizations
+pip install uniface
+
+# Linux/Windows with NVIDIA GPU
+pip install uniface[gpu]
+
+# CPU-only (all platforms)
+pip install uniface
+```
+
+---
+
+## 1. Face Detection (30 seconds)
+
+Detect faces in an image:
+
+```python
+import cv2
+from uniface import RetinaFace
+
+# Load image
+image = cv2.imread("photo.jpg")
+
+# Initialize detector (models auto-download on first use)
+detector = RetinaFace()
+
+# Detect faces
+faces = detector.detect(image)
+
+# Print results
+for i, face in enumerate(faces):
+    print(f"Face {i+1}:")
+    print(f"  Confidence: {face['confidence']:.2f}")
+    print(f"  BBox: {face['bbox']}")
+    print(f"  Landmarks: {len(face['landmarks'])} points")
+```
+
+**Output:**
+
+```
+Face 1:
+  Confidence: 0.99
+  BBox: [120.5, 85.3, 245.8, 210.6]
+  Landmarks: 5 points
+```
+
+---
+
+## 2. Visualize Detections (1 minute)
+
+Draw bounding boxes and landmarks:
+
+```python
+import cv2
+from uniface import RetinaFace
+from uniface.visualization import draw_detections
+
+# Detect faces
+detector = RetinaFace()
+image = cv2.imread("photo.jpg")
+faces = detector.detect(image)
+
+# Extract visualization data
+bboxes = [f['bbox'] for f in faces]
+scores = [f['confidence'] for f in faces]
+landmarks = [f['landmarks'] for f in faces]
+
+# Draw on image
+draw_detections(
+    image=image,
+    bboxes=bboxes,
+    scores=scores,
+    landmarks=landmarks,
+    vis_threshold=0.6,
+)
+
+# Save result
+cv2.imwrite("output.jpg", image)
+print("Saved output.jpg")
+```
+
+---
+
+## 3. Face Recognition (2 minutes)
+
+Compare two faces:
+
+```python
+import cv2
+import numpy as np
+from uniface import RetinaFace, ArcFace
+
+# Initialize models
+detector = RetinaFace()
+recognizer = ArcFace()
+
+# Load two images
+image1 = cv2.imread("person1.jpg")
+image2 = cv2.imread("person2.jpg")
+
+# Detect faces
+faces1 = detector.detect(image1)
+faces2 = detector.detect(image2)
+
+if faces1 and faces2:
+    # Extract embeddings
+    emb1 = recognizer.get_normalized_embedding(image1, faces1[0]['landmarks'])
+    emb2 = recognizer.get_normalized_embedding(image2, faces2[0]['landmarks'])
+
+    # Compute similarity (cosine similarity)
+    similarity = np.dot(emb1, emb2.T)[0][0]
+
+    # Interpret result
+    if similarity > 0.6:
+        print(f"Same person (similarity: {similarity:.3f})")
+    else:
+        print(f"Different people (similarity: {similarity:.3f})")
+else:
+    print("No faces detected")
+```
+
+**Similarity thresholds:**
+
+- `> 0.6`: Same person (high confidence)
+- `0.4 - 0.6`: Uncertain (manual review)
+- `< 0.4`: Different people
+
+---
+
+## 4. Webcam Demo (2 minutes)
+
+Real-time face detection:
+
+```python
+import cv2
+from uniface import RetinaFace
+from uniface.visualization import draw_detections
+
+detector = RetinaFace()
+cap = cv2.VideoCapture(0)
+
+print("Press 'q' to quit")
+
+while True:
+    ret, frame = cap.read()
+    if not ret:
+        break
+
+    # Detect faces
+    faces = detector.detect(frame)
+
+    # Draw results
+    bboxes = [f['bbox'] for f in faces]
+    scores = [f['confidence'] for f in faces]
+    landmarks = [f['landmarks'] for f in faces]
+    draw_detections(
+        image=frame,
+        bboxes=bboxes,
+        scores=scores,
+        landmarks=landmarks,
+    )
+
+    # Show frame
+    cv2.imshow("UniFace - Press 'q' to quit", frame)
+
+    if cv2.waitKey(1) & 0xFF == ord('q'):
+        break
+
+cap.release()
+cv2.destroyAllWindows()
+```
+
+---
+
+## 5. Age & Gender Detection (2 minutes)
+
+Detect age and gender:
+
+```python
+import cv2
+from uniface import RetinaFace, AgeGender
+
+# Initialize models
+detector = RetinaFace()
+age_gender = AgeGender()
+
+# Load image
+image = cv2.imread("photo.jpg")
+faces = detector.detect(image)
+
+# Predict attributes
+for i, face in enumerate(faces):
+    gender, age = age_gender.predict(image, face['bbox'])
+    gender_str = 'Female' if gender == 0 else 'Male'
+    print(f"Face {i+1}: {gender_str}, {age} years old")
+```
+
+**Output:**
+
+```
+Face 1: Male, 32 years old
+Face 2: Female, 28 years old
+```
+
+---
+
+## 6. Facial Landmarks (2 minutes)
+
+Detect 106 facial landmarks:
+
+```python
+import cv2
+from uniface import RetinaFace, Landmark106
+
+# Initialize models
+detector = RetinaFace()
+landmarker = Landmark106()
+
+# Detect face and landmarks
+image = cv2.imread("photo.jpg")
+faces = detector.detect(image)
+
+if faces:
+    landmarks = landmarker.get_landmarks(image, faces[0]['bbox'])
+    print(f"Detected {len(landmarks)} landmarks")
+
+    # Draw landmarks
+    for x, y in landmarks.astype(int):
+        cv2.circle(image, (x, y), 2, (0, 255, 0), -1)
+
+    cv2.imwrite("landmarks.jpg", image)
+```
+
+---
+
+## 7. Gaze Estimation (2 minutes)
+
+Estimate where a person is looking:
+
+```python
+import cv2
+import numpy as np
+from uniface import RetinaFace, MobileGaze
+from uniface.visualization import draw_gaze
+
+# Initialize models
+detector = RetinaFace()
+gaze_estimator = MobileGaze()
+
+# Load image
+image = cv2.imread("photo.jpg")
+faces = detector.detect(image)
+
+# Estimate gaze for each face
+for i, face in enumerate(faces):
+    bbox = face['bbox']
+    x1, y1, x2, y2 = map(int, bbox[:4])
+    face_crop = image[y1:y2, x1:x2]
+
+    if face_crop.size > 0:
+        pitch, yaw = gaze_estimator.estimate(face_crop)
+        print(f"Face {i+1}: pitch={np.degrees(pitch):.1f}°, yaw={np.degrees(yaw):.1f}°")
+
+        # Draw gaze direction
+        draw_gaze(image, bbox, pitch, yaw)
+
+cv2.imwrite("gaze_output.jpg", image)
+```
+
+**Output:**
+
+```
+Face 1: pitch=5.2°, yaw=-12.3°
+Face 2: pitch=-8.1°, yaw=15.7°
+```
+
+---
+
+## 8. Face Parsing (2 minutes)
+
+Segment face into semantic components (skin, eyes, nose, mouth, hair, etc.):
+
+```python
+import cv2
+import numpy as np
+from uniface.parsing import BiSeNet
+from uniface.visualization import vis_parsing_maps
+
+# Initialize parser
+parser = BiSeNet()  # Uses ResNet18 by default
+
+# Load face image (already cropped)
+face_image = cv2.imread("face.jpg")
+
+# Parse face into 19 components
+mask = parser.parse(face_image)
+
+# Visualize with overlay
+face_rgb = cv2.cvtColor(face_image, cv2.COLOR_BGR2RGB)
+vis_result = vis_parsing_maps(face_rgb, mask, save_image=False)
+
+# Convert back to BGR for saving
+vis_bgr = cv2.cvtColor(vis_result, cv2.COLOR_RGB2BGR)
+cv2.imwrite("parsed_face.jpg", vis_bgr)
+
+print(f"Detected {len(np.unique(mask))} facial components")
+```
+
+**Output:**
+
+```
+Detected 12 facial components
+```
+
+**19 Facial Component Classes:**
+- Background, Skin, Eyebrows (L/R), Eyes (L/R), Eye Glasses
+- Ears (L/R), Ear Ring, Nose, Mouth, Lips (Upper/Lower)
+- Neck, Neck Lace, Cloth, Hair, Hat
+
+---
+
+## 9. Batch Processing (3 minutes)
+
+Process multiple images:
+
+```python
+import cv2
+from pathlib import Path
+from uniface import RetinaFace
+
+detector = RetinaFace()
+
+# Process all images in a folder
+image_dir = Path("images/")
+output_dir = Path("output/")
+output_dir.mkdir(exist_ok=True)
+
+for image_path in image_dir.glob("*.jpg"):
+    print(f"Processing {image_path.name}...")
+
+    image = cv2.imread(str(image_path))
+    faces = detector.detect(image)
+
+    print(f"  Found {len(faces)} face(s)")
+
+    # Save results
+    output_path = output_dir / image_path.name
+    # ... draw and save ...
+
+print("Done!")
+```
+
+---
+
+## 10. Model Selection
+
+Choose the right model for your use case:
+
+### Detection Models
+
+```python
+from uniface.detection import RetinaFace, SCRFD, YOLOv5Face
+from uniface.constants import RetinaFaceWeights, SCRFDWeights, YOLOv5FaceWeights
+
+# Fast detection (mobile/edge devices)
+detector = RetinaFace(
+    model_name=RetinaFaceWeights.MNET_025,
+    conf_thresh=0.7
+)
+
+# Balanced (recommended)
+detector = RetinaFace(
+    model_name=RetinaFaceWeights.MNET_V2
+)
+
+# Real-time with high accuracy
+detector = YOLOv5Face(
+    model_name=YOLOv5FaceWeights.YOLOV5S,
+    conf_thresh=0.6,
+    nms_thresh=0.5
+)
+
+# High accuracy (server/GPU)
+detector = SCRFD(
+    model_name=SCRFDWeights.SCRFD_10G_KPS,
+    conf_thresh=0.5
+)
+```
+
+### Recognition Models
+
+```python
+from uniface import ArcFace, MobileFace, SphereFace
+from uniface.constants import MobileFaceWeights, SphereFaceWeights
+
+# ArcFace (recommended for most use cases)
+recognizer = ArcFace()  # Best accuracy
+
+# MobileFace (lightweight for mobile/edge)
+recognizer = MobileFace(model_name=MobileFaceWeights.MNET_V2)  # Fast, small size
+
+# SphereFace (angular margin approach)
+recognizer = SphereFace(model_name=SphereFaceWeights.SPHERE20)  # Alternative method
+```
+
+### Gaze Estimation Models
+
+```python
+from uniface import MobileGaze
+from uniface.constants import GazeWeights
+
+# Default (recommended)
+gaze_estimator = MobileGaze()  # Uses RESNET34
+
+# Lightweight (mobile/edge devices)
+gaze_estimator = MobileGaze(model_name=GazeWeights.MOBILEONE_S0)
+
+# High accuracy
+gaze_estimator = MobileGaze(model_name=GazeWeights.RESNET50)
+```
+
+### Face Parsing Models
+
+```python
+from uniface.parsing import BiSeNet
+from uniface.constants import ParsingWeights
+
+# Default (recommended, 50.7 MB)
+parser = BiSeNet()  # Uses RESNET18
+
+# Higher accuracy (89.2 MB)
+parser = BiSeNet(model_name=ParsingWeights.RESNET34)
+```
+
+---
+
+## Common Issues
+
+### 1. Models Not Downloading
+
+```python
+# Manually download a model
+from uniface.model_store import verify_model_weights
+from uniface.constants import RetinaFaceWeights
+
+model_path = verify_model_weights(RetinaFaceWeights.MNET_V2)
+print(f"Model downloaded to: {model_path}")
+```
+
+### 2. Check Hardware Acceleration
+
+```python
+import onnxruntime as ort
+print("Available providers:", ort.get_available_providers())
+
+# macOS M-series should show: ['CoreMLExecutionProvider', ...]
+# NVIDIA GPU should show: ['CUDAExecutionProvider', ...]
+```
+
+### 3. Slow Performance on Mac
+
+The standard installation includes ARM64 optimizations for Apple Silicon. If performance is slow, verify you're using the ARM64 build of Python:
+
+```bash
+python -c "import platform; print(platform.machine())"
+# Should show: arm64 (not x86_64)
+```
+
+### 4. Import Errors
+
+```python
+# Correct imports
+from uniface.detection import RetinaFace
+from uniface.recognition import ArcFace
+from uniface.landmark import Landmark106
+
+# Wrong imports
+from uniface import retinaface  # Module, not class
+```
+
+---
+
+## Next Steps
+
+### Jupyter Notebook Examples
+
+Explore interactive examples for common tasks:
+
+| Example | Description | Notebook |
+|---------|-------------|----------|
+| **Face Detection** | Detect faces and facial landmarks | [face_detection.ipynb](examples/face_detection.ipynb) |
+| **Face Alignment** | Align and crop faces for recognition | [face_alignment.ipynb](examples/face_alignment.ipynb) |
+| **Face Recognition** | Extract face embeddings and compare faces | [face_analyzer.ipynb](examples/face_analyzer.ipynb) |
+| **Face Verification** | Compare two faces to verify identity | [face_verification.ipynb](examples/face_verification.ipynb) |
+| **Face Search** | Find a person in a group photo | [face_search.ipynb](examples/face_search.ipynb) |
+| **Face Parsing** | Segment face into semantic components | [face_parsing.ipynb](examples/face_parsing.ipynb) |
+| **Gaze Estimation** | Estimate gaze direction | [gaze_estimation.ipynb](examples/gaze_estimation.ipynb) |
+
+### Additional Resources
+
+- **Model Benchmarks**: See [MODELS.md](MODELS.md) for performance comparisons
+- **Full Documentation**: Read [README.md](README.md) for complete API reference
+
+---
+
+## References
+
+- **RetinaFace Training**: [yakhyo/retinaface-pytorch](https://github.com/yakhyo/retinaface-pytorch)
+- **YOLOv5-Face ONNX**: [yakhyo/yolov5-face-onnx-inference](https://github.com/yakhyo/yolov5-face-onnx-inference)
+- **Face Recognition Training**: [yakhyo/face-recognition](https://github.com/yakhyo/face-recognition)
+- **Gaze Estimation Training**: [yakhyo/gaze-estimation](https://github.com/yakhyo/gaze-estimation)
+- **Face Parsing Training**: [yakhyo/face-parsing](https://github.com/yakhyo/face-parsing)
+- **InsightFace**: [deepinsight/insightface](https://github.com/deepinsight/insightface)

README.md

@@ -1,239 +1,575 @@
 # UniFace: All-in-One Face Analysis Library
 
-<div align="center">
-
 [![License](https://img.shields.io/badge/License-MIT-blue.svg)](https://opensource.org/licenses/MIT)
-![Python](https://img.shields.io/badge/Python-3.8%2B-blue)
-[![PyPI Version](https://img.shields.io/pypi/v/uniface.svg)](https://pypi.org/project/uniface/)
-[![Build Status](https://github.com/yakhyo/uniface/actions/workflows/build.yml/badge.svg)](https://github.com/yakhyo/uniface/actions)
+[![Python](https://img.shields.io/badge/Python-3.10%2B-blue)](https://www.python.org/)
+[![PyPI](https://img.shields.io/pypi/v/uniface.svg)](https://pypi.org/project/uniface/)
+[![CI](https://github.com/yakhyo/uniface/actions/workflows/ci.yml/badge.svg)](https://github.com/yakhyo/uniface/actions)
 [![Downloads](https://pepy.tech/badge/uniface)](https://pepy.tech/project/uniface)
-[![Code Style: PEP8](https://img.shields.io/badge/code%20style-PEP8-green.svg)](https://www.python.org/dev/peps/pep-0008/)
-[![GitHub Release Downloads](https://img.shields.io/github/downloads/yakhyo/uniface/total.svg?label=Model%20Downloads)](https://github.com/yakhyo/uniface/releases)
+[![DeepWiki](https://img.shields.io/badge/DeepWiki-yakhyo%2Funiface-blue.svg?logo=data:image/png;base64,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)](https://deepwiki.com/yakhyo/uniface)
 
+<div align="center">
+    <img src=".github/logos/logo_web.webp" width=75%>
 </div>
 
-**uniface** is a lightweight face detection library designed for high-performance face localization and landmark detection. The library supports ONNX models and provides utilities for bounding box visualization and landmark plotting. To train RetinaFace model, see https://github.com/yakhyo/retinaface-pytorch.
+**UniFace** is a lightweight, production-ready face analysis library built on ONNX Runtime. It provides high-performance face detection, recognition, landmark detection, face parsing, gaze estimation, and attribute analysis with hardware acceleration support across platforms.
 
 ---
 
 ## Features
-- [ ] Age and gender detection (Planned).
-- [ ] Face recognition (Planned).
-- [x] High-speed face detection using ONNX models (Added: 2024-11-20).
-- [x] Accurate facial landmark localization (e.g., eyes, nose, and mouth) (Added: 2024-11-20).
-- [x] Easy-to-use API for inference and visualization (Added: 2024-11-20).
+
+- **High-Speed Face Detection**: ONNX-optimized RetinaFace, SCRFD, and YOLOv5-Face models
+- **Facial Landmark Detection**: Accurate 106-point landmark localization
+- **Face Recognition**: ArcFace, MobileFace, and SphereFace embeddings
+- **Face Parsing**: BiSeNet-based semantic segmentation with 19 facial component classes
+- **Gaze Estimation**: Real-time gaze direction prediction with MobileGaze
+- **Attribute Analysis**: Age, gender, and emotion detection
+- **Face Alignment**: Precise alignment for downstream tasks
+- **Hardware Acceleration**: ARM64 optimizations (Apple Silicon), CUDA (NVIDIA), CPU fallback
+- **Simple API**: Intuitive factory functions and clean interfaces
+- **Production-Ready**: Type hints, comprehensive logging, PEP8 compliant
 
 ---
 
 ## Installation
 
-### Using pip
+### Quick Install (All Platforms)
 
 ```bash
 pip install uniface
 ```
 
-### Local installation using pip
+### Platform-Specific Installation
 
-**Clone the repository**
+#### macOS (Apple Silicon - M1/M2/M3/M4)
+
+For Apple Silicon Macs, the standard installation automatically includes optimized ARM64 support:
+
+```bash
+pip install uniface
+```
+
+The base `onnxruntime` package (included with uniface) has native Apple Silicon support with ARM64 optimizations built-in since version 1.13+.
+
+#### Linux/Windows with NVIDIA GPU
+
+For CUDA acceleration on NVIDIA GPUs:
+
+```bash
+pip install uniface[gpu]
+```
+
+**Requirements:**
+
+- CUDA 11.x or 12.x
+- cuDNN 8.x
+- See [ONNX Runtime GPU requirements](https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html)
+
+#### CPU-Only (All Platforms)
+
+```bash
+pip install uniface
+```
+
+### Install from Source
 
 ```bash
 git clone https://github.com/yakhyo/uniface.git
 cd uniface
-```
-
-**Install using pip**
-
-```bash
-pip install .
+pip install -e .
 ```
 
 ---
 
 ## Quick Start
 
-### Initialize the Model
+### Face Detection
 
 ```python
+import cv2
 from uniface import RetinaFace
 
-# Initialize the RetinaFace model
-uniface_inference = RetinaFace(
-    model="retinaface_mnet_v2",  # Model name
-    conf_thresh=0.5,             # Confidence threshold
-    pre_nms_topk=5000,           # Pre-NMS Top-K detections
-    nms_thresh=0.4,              # NMS IoU threshold
-    post_nms_topk=750            # Post-NMS Top-K detections
+# Initialize detector
+detector = RetinaFace()
+
+# Load image
+image = cv2.imread("image.jpg")
+
+# Detect faces
+faces = detector.detect(image)
+
+# Process results
+for face in faces:
+    bbox = face['bbox']  # [x1, y1, x2, y2]
+    confidence = face['confidence']
+    landmarks = face['landmarks']  # 5-point landmarks
+    print(f"Face detected with confidence: {confidence:.2f}")
+```
+
+### Face Recognition
+
+```python
+from uniface import ArcFace, RetinaFace
+from uniface import compute_similarity
+
+# Initialize models
+detector = RetinaFace()
+recognizer = ArcFace()
+
+# Detect and extract embeddings
+faces1 = detector.detect(image1)
+faces2 = detector.detect(image2)
+
+embedding1 = recognizer.get_normalized_embedding(image1, faces1[0]['landmarks'])
+embedding2 = recognizer.get_normalized_embedding(image2, faces2[0]['landmarks'])
+
+# Compare faces
+similarity = compute_similarity(embedding1, embedding2)
+print(f"Similarity: {similarity:.4f}")
+```
+
+### Facial Landmarks
+
+```python
+from uniface import RetinaFace, Landmark106
+
+detector = RetinaFace()
+landmarker = Landmark106()
+
+faces = detector.detect(image)
+landmarks = landmarker.get_landmarks(image, faces[0]['bbox'])
+# Returns 106 (x, y) landmark points
+```
+
+### Age & Gender Detection
+
+```python
+from uniface import RetinaFace, AgeGender
+
+detector = RetinaFace()
+age_gender = AgeGender()
+
+faces = detector.detect(image)
+gender, age = age_gender.predict(image, faces[0]['bbox'])
+gender_str = 'Female' if gender == 0 else 'Male'
+print(f"{gender_str}, {age} years old")
+```
+
+### Gaze Estimation
+
+```python
+from uniface import RetinaFace, MobileGaze
+from uniface.visualization import draw_gaze
+import numpy as np
+
+detector = RetinaFace()
+gaze_estimator = MobileGaze()
+
+faces = detector.detect(image)
+for face in faces:
+    bbox = face['bbox']
+    x1, y1, x2, y2 = map(int, bbox[:4])
+    face_crop = image[y1:y2, x1:x2]
+
+    pitch, yaw = gaze_estimator.estimate(face_crop)
+    print(f"Gaze: pitch={np.degrees(pitch):.1f}°, yaw={np.degrees(yaw):.1f}°")
+
+    # Visualize
+    draw_gaze(image, bbox, pitch, yaw)
+```
+
+### Face Parsing
+
+```python
+from uniface.parsing import BiSeNet
+from uniface.visualization import vis_parsing_maps
+
+# Initialize parser
+parser = BiSeNet()  # Uses ResNet18 by default
+
+# Parse face image (already cropped)
+mask = parser.parse(face_image)
+
+# Visualize with overlay
+import cv2
+face_rgb = cv2.cvtColor(face_image, cv2.COLOR_BGR2RGB)
+vis_result = vis_parsing_maps(face_rgb, mask, save_image=False)
+
+# mask contains 19 classes: skin, eyes, nose, mouth, hair, etc.
+print(f"Unique classes: {len(np.unique(mask))}")
+```
+
+---
+
+## Documentation
+
+- [**QUICKSTART.md**](QUICKSTART.md) - 5-minute getting started guide
+- [**MODELS.md**](MODELS.md) - Model zoo, benchmarks, and selection guide
+- [**Examples**](examples/) - Jupyter notebooks with detailed examples
+
+---
+
+## API Overview
+
+### Factory Functions (Recommended)
+
+```python
+from uniface.detection import RetinaFace, SCRFD
+from uniface.recognition import ArcFace
+from uniface.landmark import Landmark106
+
+from uniface.constants import SCRFDWeights
+
+# Create detector with default settings
+detector = RetinaFace()
+
+# Create with custom config
+detector = SCRFD(
+    model_name=SCRFDWeights.SCRFD_10G_KPS, # SCRFDWeights.SCRFD_500M_KPS
+    conf_thresh=0.4,
+    input_size=(640, 640)
 )
+# Or with defaults settings: detector = SCRFD()
+
+# Recognition and landmarks
+recognizer = ArcFace()
+landmarker = Landmark106()
 ```
 
-### Run Inference
+### Direct Model Instantiation
 
-Inference on image:
+```python
+from uniface import RetinaFace, SCRFD, YOLOv5Face, ArcFace, MobileFace, SphereFace
+from uniface.constants import RetinaFaceWeights, YOLOv5FaceWeights
+
+# Detection
+detector = RetinaFace(
+    model_name=RetinaFaceWeights.MNET_V2,
+    conf_thresh=0.5,
+    nms_thresh=0.4
+)
+# Or detector = RetinaFace()
+
+# YOLOv5-Face detection
+detector = YOLOv5Face(
+    model_name=YOLOv5FaceWeights.YOLOV5S,
+    conf_thresh=0.6,
+    nms_thresh=0.5
+)
+# Or detector = YOLOv5Face
+
+# Recognition
+recognizer = ArcFace()  # Uses default weights
+recognizer = MobileFace()  # Lightweight alternative
+recognizer = SphereFace()  # Angular softmax alternative
+```
+
+### High-Level Detection API
+
+```python
+from uniface import detect_faces
+
+# One-line face detection
+faces = detect_faces(image, method='retinaface', conf_thresh=0.8)  # methods: retinaface, scrfd, yolov5face
+```
+
+### Key Parameters (quick reference)
+
+**Detection**
+
+| Class          | Key params (defaults)                                                                                                                | Notes                                          |
+| -------------- | ------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------- |
+| `RetinaFace` | `model_name=RetinaFaceWeights.MNET_V2`, `conf_thresh=0.5`, `nms_thresh=0.4`, `input_size=(640, 640)`, `dynamic_size=False` | Supports 5-point landmarks                     |
+| `SCRFD`      | `model_name=SCRFDWeights.SCRFD_10G_KPS`, `conf_thresh=0.5`, `nms_thresh=0.4`, `input_size=(640, 640)`                        | Supports 5-point landmarks                     |
+| `YOLOv5Face` | `model_name=YOLOv5FaceWeights.YOLOV5S`, `conf_thresh=0.6`, `nms_thresh=0.5`, `input_size=640` (fixed)                        | Supports 5-point landmarks; models: YOLOV5N/S/M; `input_size` must be 640 |
+
+**Recognition**
+
+| Class          | Key params (defaults)                     | Notes                                 |
+| -------------- | ----------------------------------------- | ------------------------------------- |
+| `ArcFace`    | `model_name=ArcFaceWeights.MNET`        | Returns 512-dim normalized embeddings |
+| `MobileFace` | `model_name=MobileFaceWeights.MNET_V2`  | Lightweight embeddings                |
+| `SphereFace` | `model_name=SphereFaceWeights.SPHERE20` | Angular softmax variant               |
+
+**Landmark & Attributes**
+
+| Class           | Key params (defaults)                                                 | Notes                                   |
+| --------------- | --------------------------------------------------------------------- | --------------------------------------- |
+| `Landmark106` | No required params                                                    | 106-point landmarks                     |
+| `AgeGender`   | `model_name=AgeGenderWeights.DEFAULT`; `input_size` auto-detected | Requires bbox; ONNXRuntime              |
+| `Emotion`     | `model_weights=DDAMFNWeights.AFFECNET7`, `input_size=(112, 112)`  | Requires 5-point landmarks; TorchScript |
+
+**Gaze Estimation**
+
+| Class         | Key params (defaults)                      | Notes                                |
+| ------------- | ------------------------------------------ | ------------------------------------ |
+| `MobileGaze` | `model_name=GazeWeights.RESNET34`       | Returns (pitch, yaw) angles in radians; trained on Gaze360 |
+
+**Face Parsing**
+
+| Class      | Key params (defaults)                    | Notes                                |
+| ---------- | ---------------------------------------- | ------------------------------------ |
+| `BiSeNet` | `model_name=ParsingWeights.RESNET18`, `input_size=(512, 512)` | 19 facial component classes; BiSeNet architecture with ResNet backbone |
+
+---
+
+## Model Performance
+
+### Face Detection (WIDER FACE Dataset)
+
+| Model              | Easy   | Medium | Hard   | Use Case               |
+| ------------------ | ------ | ------ | ------ | ---------------------- |
+| retinaface_mnet025 | 88.48% | 87.02% | 80.61% | Mobile/Edge devices    |
+| retinaface_mnet_v2 | 91.70% | 91.03% | 86.60% | Balanced (recommended) |
+| retinaface_r34     | 94.16% | 93.12% | 88.90% | High accuracy          |
+| scrfd_500m         | 90.57% | 88.12% | 68.51% | Real-time applications |
+| scrfd_10g          | 95.16% | 93.87% | 83.05% | Best accuracy/speed    |
+| yolov5n_face       | 93.61% | 91.52% | 80.53% | Lightweight/Mobile     |
+| yolov5s_face       | 94.33% | 92.61% | 83.15% | Real-time + accuracy   |
+| yolov5m_face       | 95.30% | 93.76% | 85.28% | High accuracy          |
+
+_Accuracy values from original papers: [RetinaFace](https://arxiv.org/abs/1905.00641), [SCRFD](https://arxiv.org/abs/2105.04714), [YOLOv5-Face](https://arxiv.org/abs/2105.12931)_
+
+**Benchmark on your hardware:**
+
+```bash
+python scripts/run_detection.py --image assets/test.jpg --iterations 100
+```
+
+See [MODELS.md](MODELS.md) for detailed model information and selection guide.
+
+<div align="center">
+    <img src="assets/test_result.png">
+</div>
+
+---
+
+## Examples
+
+### Jupyter Notebooks
+
+Interactive examples covering common face analysis tasks:
+
+| Example | Description | Notebook |
+|---------|-------------|----------|
+| **Face Detection** | Detect faces and facial landmarks | [face_detection.ipynb](examples/face_detection.ipynb) |
+| **Face Alignment** | Align and crop faces for recognition | [face_alignment.ipynb](examples/face_alignment.ipynb) |
+| **Face Recognition** | Extract face embeddings and compare faces | [face_analyzer.ipynb](examples/face_analyzer.ipynb) |
+| **Face Verification** | Compare two faces to verify identity | [face_verification.ipynb](examples/face_verification.ipynb) |
+| **Face Search** | Find a person in a group photo | [face_search.ipynb](examples/face_search.ipynb) |
+| **Face Parsing** | Segment face into semantic components | [face_parsing.ipynb](examples/face_parsing.ipynb) |
+| **Gaze Estimation** | Estimate gaze direction from face images | [gaze_estimation.ipynb](examples/gaze_estimation.ipynb) |
+
+### Webcam Face Detection
 
 ```python
 import cv2
+from uniface import RetinaFace
 from uniface.visualization import draw_detections
 
-# Load an image
-image_path = "assets/test.jpg"
-original_image = cv2.imread(image_path)
-
-# Perform inference
-boxes, landmarks = uniface_inference.detect(original_image)
-
-# Visualize results
-draw_detections(original_image, (boxes, landmarks), vis_threshold=0.6)
-
-# Save the output image
-output_path = "output.jpg"
-cv2.imwrite(output_path, original_image)
-print(f"Saved output image to {output_path}")
-```
-
-Inference on video:
-
-```python
-import cv2
-from uniface.visualization import draw_detections
-
-# Initialize the webcam
+detector = RetinaFace()
 cap = cv2.VideoCapture(0)
 
-if not cap.isOpened():
-    print("Error: Unable to access the webcam.")
-    exit()
-
 while True:
-    # Capture a frame from the webcam
     ret, frame = cap.read()
     if not ret:
-        print("Error: Failed to read frame.")
         break
 
-    # Perform inference
-    boxes, landmarks = uniface_inference.detect(frame)
+    faces = detector.detect(frame)
 
-    # Draw detections on the frame
-    draw_detections(frame, (boxes, landmarks), vis_threshold=0.6)
+    # Extract data for visualization
+    bboxes = [f['bbox'] for f in faces]
+    scores = [f['confidence'] for f in faces]
+    landmarks = [f['landmarks'] for f in faces]
 
-    # Display the output
-    cv2.imshow("Webcam Inference", frame)
+    draw_detections(
+        image=frame,
+        bboxes=bboxes,
+        scores=scores,
+        landmarks=landmarks,
+        vis_threshold=0.6,
+    )
 
-    # Exit if 'q' is pressed
+    cv2.imshow("Face Detection", frame)
     if cv2.waitKey(1) & 0xFF == ord('q'):
         break
 
-# Release the webcam and close all OpenCV windows
 cap.release()
 cv2.destroyAllWindows()
 ```
 
+### Face Search System
+
+```python
+import numpy as np
+from uniface import RetinaFace, ArcFace
+
+detector = RetinaFace()
+recognizer = ArcFace()
+
+# Build face database
+database = {}
+for person_id, image_path in person_images.items():
+    image = cv2.imread(image_path)
+    faces = detector.detect(image)
+    if faces:
+        embedding = recognizer.get_normalized_embedding(
+            image, faces[0]['landmarks']
+        )
+        database[person_id] = embedding
+
+# Search for a face
+query_image = cv2.imread("query.jpg")
+query_faces = detector.detect(query_image)
+if query_faces:
+    query_embedding = recognizer.get_normalized_embedding(
+        query_image, query_faces[0]['landmarks']
+    )
+
+    # Find best match
+    best_match = None
+    best_similarity = -1
+
+    for person_id, db_embedding in database.items():
+        similarity = np.dot(query_embedding, db_embedding.T)[0][0]
+        if similarity > best_similarity:
+            best_similarity = similarity
+            best_match = person_id
+
+    print(f"Best match: {best_match} (similarity: {best_similarity:.4f})")
+```
+
+More examples in the [examples/](examples/) directory.
+
 ---
 
-### Evaluation results of available models on WiderFace
+## Advanced Configuration
 
-| RetinaFace Models  | Easy       | Medium     | Hard       |
-| ------------------ | ---------- | ---------- | ---------- |
-| retinaface_mnet025 | 88.48%     | 87.02%     | 80.61%     |
-| retinaface_mnet050 | 89.42%     | 87.97%     | 82.40%     |
-| retinaface_mnet_v1 | 90.59%     | 89.14%     | 84.13%     |
-| retinaface_mnet_v2 | 91.70%     | 91.03%     | 86.60%     |
-| retinaface_r18     | 92.50%     | 91.02%     | 86.63%     |
-| retinaface_r34     | **94.16%** | **93.12%** | **88.90%** |
+### Custom ONNX Runtime Providers
 
-## API Reference
-
-### `RetinaFace` Class
-
-#### Initialization
 ```python
-RetinaFace(
-    model: str,
-    conf_thresh: float = 0.5,
-    pre_nms_topk: int = 5000,
-    nms_thresh: float = 0.4,
-    post_nms_topk: int = 750
+from uniface.onnx_utils import get_available_providers, create_onnx_session
+
+# Check available providers
+providers = get_available_providers()
+print(f"Available: {providers}")
+
+# Force CPU-only execution
+from uniface import RetinaFace
+detector = RetinaFace()
+# Internally uses create_onnx_session() which auto-selects best provider
+```
+
+### Model Download and Caching
+
+Models are automatically downloaded on first use and cached in `~/.uniface/models/`.
+
+```python
+from uniface.model_store import verify_model_weights
+from uniface.constants import RetinaFaceWeights
+
+# Manually download and verify a model
+model_path = verify_model_weights(
+    RetinaFaceWeights.MNET_V2,
+    root='./custom_models'  # Custom cache directory
 )
 ```
 
-**Parameters**:
-- `model` *(str)*: Name of the model to use. Supported models:
-  - `retinaface_mnet025`, `retinaface_mnet050`, `retinaface_mnet_v1`, `retinaface_mnet_v2`
-  - `retinaface_r18`, `retinaface_r34`
-- `conf_thresh` *(float, default=0.5)*: Minimum confidence score for detections.
-- `pre_nms_topk` *(int, default=5000)*: Max detections to keep before NMS.
-- `nms_thresh` *(float, default=0.4)*: IoU threshold for Non-Maximum Suppression.
-- `post_nms_topk` *(int, default=750)*: Max detections to keep after NMS.
+### Logging Configuration
 
----
-
-### `detect` Method
 ```python
-detect(
-    image: np.ndarray,
-    max_num: int = 0,
-    metric: str = "default",
-    center_weight: float = 2.0
-) -> Tuple[np.ndarray, np.ndarray]
+from uniface import Logger
+import logging
+
+# Set logging level
+Logger.setLevel(logging.DEBUG)  # DEBUG, INFO, WARNING, ERROR
+
+# Disable logging
+Logger.setLevel(logging.CRITICAL)
 ```
 
-**Description**:
-Detects faces in the given image and returns bounding boxes and landmarks.
-
-**Parameters**:
-- `image` *(np.ndarray)*: Input image in BGR format.
-- `max_num` *(int, default=0)*: Maximum number of faces to return. `0` means return all.
-- `metric` *(str, default="default")*: Metric for prioritizing detections:
-  - `"default"`: Prioritize detections closer to the image center.
-  - `"max"`: Prioritize larger bounding box areas.
-- `center_weight` *(float, default=2.0)*: Weight for prioritizing center-aligned faces.
-
-**Returns**:
-- `bounding_boxes` *(np.ndarray)*: Array of detections as `[x_min, y_min, x_max, y_max, confidence]`.
-- `landmarks` *(np.ndarray)*: Array of landmarks as `[(x1, y1), ..., (x5, y5)]`.
-
 ---
 
-### Visualization Utilities
+## Testing
 
-#### `draw_detections`
-```python
-draw_detections(
-    image: np.ndarray,
-    detections: Tuple[np.ndarray, np.ndarray],
-    vis_threshold: float
-) -> None
+```bash
+# Run all tests
+pytest
+
+# Run with coverage
+pytest --cov=uniface --cov-report=html
+
+# Run specific test file
+pytest tests/test_retinaface.py -v
 ```
 
-**Description**:
-Draws bounding boxes and landmarks on the given image.
+---
 
-**Parameters**:
-- `image` *(np.ndarray)*: The input image in BGR format.
-- `detections` *(Tuple[np.ndarray, np.ndarray])*: A tuple of bounding boxes and landmarks.
-- `vis_threshold` *(float)*: Minimum confidence score for visualization.
+## Development
+
+### Setup Development Environment
+
+```bash
+git clone https://github.com/yakhyo/uniface.git
+cd uniface
+
+# Install in editable mode with dev dependencies
+pip install -e ".[dev]"
+
+# Run tests
+pytest
+```
+
+### Code Formatting
+
+This project uses [Ruff](https://docs.astral.sh/ruff/) for linting and formatting.
+
+```bash
+# Format code
+ruff format .
+
+# Check for linting errors
+ruff check .
+
+# Auto-fix linting errors
+ruff check . --fix
+```
+
+Ruff configuration is in `pyproject.toml`. Key settings:
+
+- Line length: 120
+- Python target: 3.10+
+- Import sorting: `uniface` as first-party
+
+### Project Structure
+
+```
+uniface/
+├── uniface/
+│   ├── detection/       # Face detection models
+│   ├── recognition/     # Face recognition models
+│   ├── landmark/        # Landmark detection
+│   ├── parsing/         # Face parsing
+│   ├── gaze/            # Gaze estimation
+│   ├── attribute/       # Age, gender, emotion
+│   ├── onnx_utils.py    # ONNX Runtime utilities
+│   ├── model_store.py   # Model download & caching
+│   └── visualization.py # Drawing utilities
+├── tests/               # Unit tests
+├── examples/            # Example notebooks
+└── scripts/             # Utility scripts
+```
 
 ---
 
+## References
+
+- **RetinaFace Training**: [yakhyo/retinaface-pytorch](https://github.com/yakhyo/retinaface-pytorch) - PyTorch implementation and training code
+- **YOLOv5-Face ONNX**: [yakhyo/yolov5-face-onnx-inference](https://github.com/yakhyo/yolov5-face-onnx-inference) - ONNX inference implementation
+- **Face Recognition Training**: [yakhyo/face-recognition](https://github.com/yakhyo/face-recognition) - ArcFace, MobileFace, SphereFace training code
+- **Face Parsing Training**: [yakhyo/face-parsing](https://github.com/yakhyo/face-parsing) - BiSeNet face parsing training code and pretrained weights
+- **Gaze Estimation Training**: [yakhyo/gaze-estimation](https://github.com/yakhyo/gaze-estimation) - MobileGaze training code and pretrained weights
+- **InsightFace**: [deepinsight/insightface](https://github.com/deepinsight/insightface) - Model architectures and pretrained weights
+
 ## Contributing
 
-We welcome contributions to enhance the library! Feel free to:
-
-- Submit bug reports or feature requests.
-- Fork the repository and create a pull request.
-
----
-
-## License
-
-This project is licensed under the MIT License. See the [LICENSE](LICENSE) file for details.
-
----
-
-## Acknowledgments
-
-- Based on the RetinaFace model for face detection ([https://github.com/yakhyo/retinaface-pytorch](https://github.com/yakhyo/retinaface-pytorch)).
-- Inspired by InsightFace and other face detection projects.
-
----
+Contributions are welcome! Please open an issue or submit a pull request on [GitHub](https://github.com/yakhyo/uniface).

pyproject.toml

@@ -0,0 +1,102 @@
+[project]
+name = "uniface"
+version = "1.5.0"
+description = "UniFace: A Comprehensive Library for Face Detection, Recognition, Landmark Analysis, Face Parsing, Gaze Estimation, Age, and Gender Detection"
+readme = "README.md"
+license = { text = "MIT" }
+authors = [{ name = "Yakhyokhuja Valikhujaev", email = "yakhyo9696@gmail.com" }]
+maintainers = [
+    { name = "Yakhyokhuja Valikhujaev", email = "yakhyo9696@gmail.com" },
+]
+
+requires-python = ">=3.10,<3.14"
+keywords = [
+    "face-detection",
+    "face-recognition",
+    "facial-landmarks",
+    "face-parsing",
+    "face-segmentation",
+    "gaze-estimation",
+    "age-detection",
+    "gender-detection",
+    "computer-vision",
+    "deep-learning",
+    "onnx",
+    "onnxruntime",
+    "face-analysis",
+    "bisenet",
+]
+
+classifiers = [
+    "Development Status :: 4 - Beta",
+    "Intended Audience :: Developers",
+    "Intended Audience :: Science/Research",
+    "License :: OSI Approved :: MIT License",
+    "Operating System :: OS Independent",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
+]
+
+dependencies = [
+    "numpy>=1.21.0",
+    "opencv-python>=4.5.0",
+    "onnx>=1.12.0",
+    "onnxruntime>=1.16.0",
+    "scikit-image>=0.19.0",
+    "requests>=2.28.0",
+    "tqdm>=4.64.0",
+]
+
+[project.optional-dependencies]
+dev = ["pytest>=7.0.0", "ruff>=0.4.0"]
+gpu = ["onnxruntime-gpu>=1.16.0"]
+
+[project.urls]
+Homepage = "https://github.com/yakhyo/uniface"
+Repository = "https://github.com/yakhyo/uniface"
+Documentation = "https://github.com/yakhyo/uniface/blob/main/README.md"
+"Quick Start" = "https://github.com/yakhyo/uniface/blob/main/QUICKSTART.md"
+"Model Zoo" = "https://github.com/yakhyo/uniface/blob/main/MODELS.md"
+
+[build-system]
+requires = ["setuptools>=64", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[tool.setuptools]
+packages = { find = { where = ["."], include = ["uniface*"] } }
+
+[tool.setuptools.package-data]
+uniface = ["py.typed"]
+
+[tool.ruff]
+line-length = 120
+target-version = "py310"
+exclude = [
+    ".git",
+    ".ruff_cache",
+    "__pycache__",
+    "build",
+    "dist",
+    "*.egg-info",
+    ".venv",
+    "venv",
+    ".pytest_cache",
+    ".mypy_cache",
+    "*.ipynb",
+]
+
+[tool.ruff.format]
+quote-style = "single"
+
+
+[tool.ruff.lint]
+select = ["E", "F", "I", "W"]
+
+[tool.ruff.lint.flake8-quotes]
+docstring-quotes = "double"
+
+[tool.ruff.lint.isort]
+known-first-party = ["uniface"]

requirements.txt

@@ -1,8 +1,8 @@
-pytest
-numpy
-opencv-python
-opencv-python-headless
-onnx
-onnxruntime
-requests
-torch
+numpy>=1.21.0
+opencv-python>=4.5.0
+onnx>=1.12.0
+onnxruntime>=1.16.0
+scikit-image>=0.19.0
+requests>=2.28.0
+pytest>=7.0.0
+tqdm>=4.64.0

scripts/README.md

@@ -0,0 +1,79 @@
+# Scripts
+
+Scripts for testing UniFace features.
+
+## Available Scripts
+
+| Script | Description |
+|--------|-------------|
+| `run_detection.py` | Face detection on image or webcam |
+| `run_age_gender.py` | Age and gender prediction |
+| `run_emotion.py` | Emotion detection (7 or 8 emotions) |
+| `run_gaze_estimation.py` | Gaze direction estimation |
+| `run_landmarks.py` | 106-point facial landmark detection |
+| `run_recognition.py` | Face embedding extraction and comparison |
+| `run_face_analyzer.py` | Complete face analysis (detection + recognition + attributes) |
+| `run_face_search.py` | Real-time face matching against reference |
+| `run_video_detection.py` | Face detection on video files |
+| `batch_process.py` | Batch process folder of images |
+| `download_model.py` | Download model weights |
+| `sha256_generate.py` | Generate SHA256 hash for model files |
+
+## Usage Examples
+
+```bash
+# Face detection
+python scripts/run_detection.py --image assets/test.jpg
+python scripts/run_detection.py --webcam
+
+# Age and gender
+python scripts/run_age_gender.py --image assets/test.jpg
+python scripts/run_age_gender.py --webcam
+
+# Emotion detection
+python scripts/run_emotion.py --image assets/test.jpg
+python scripts/run_emotion.py --webcam
+
+# Gaze estimation
+python scripts/run_gaze_estimation.py --image assets/test.jpg
+python scripts/run_gaze_estimation.py --webcam
+
+# Landmarks
+python scripts/run_landmarks.py --image assets/test.jpg
+python scripts/run_landmarks.py --webcam
+
+# Face recognition (extract embedding)
+python scripts/run_recognition.py --image assets/test.jpg
+
+# Face comparison
+python scripts/run_recognition.py --image1 face1.jpg --image2 face2.jpg
+
+# Face search (match webcam against reference)
+python scripts/run_face_search.py --image reference.jpg
+
+# Video processing
+python scripts/run_video_detection.py --input video.mp4 --output output.mp4
+
+# Batch processing
+python scripts/batch_process.py --input images/ --output results/
+
+# Download models
+python scripts/download_model.py --model-type retinaface
+python scripts/download_model.py  # downloads all
+```
+
+## Common Options
+
+| Option | Description |
+|--------|-------------|
+| `--image` | Path to input image |
+| `--webcam` | Use webcam instead of image |
+| `--method` | Choose detector: `retinaface`, `scrfd`, `yolov5face` |
+| `--threshold` | Visualization confidence threshold (default: 0.25) |
+| `--save_dir` | Output directory (default: `outputs`) |
+
+## Quick Test
+
+```bash
+python scripts/run_detection.py --image assets/test.jpg
+```

scripts/batch_process.py

@@ -0,0 +1,98 @@
+# Batch face detection on a folder of images
+# Usage: python batch_process.py --input images/ --output results/
+
+import argparse
+from pathlib import Path
+
+import cv2
+from tqdm import tqdm
+
+from uniface import SCRFD, RetinaFace
+from uniface.visualization import draw_detections
+
+
+def get_image_files(input_dir: Path, extensions: tuple) -> list:
+    files = []
+    for ext in extensions:
+        files.extend(input_dir.glob(f'*.{ext}'))
+        files.extend(input_dir.glob(f'*.{ext.upper()}'))
+    return sorted(files)
+
+
+def process_image(detector, image_path: Path, output_path: Path, threshold: float) -> int:
+    """Process single image. Returns face count or -1 on error."""
+    image = cv2.imread(str(image_path))
+    if image is None:
+        return -1
+
+    faces = detector.detect(image)
+
+    # unpack face data for visualization
+    bboxes = [f['bbox'] for f in faces]
+    scores = [f['confidence'] for f in faces]
+    landmarks = [f['landmarks'] for f in faces]
+    draw_detections(
+        image=image, bboxes=bboxes, scores=scores, landmarks=landmarks, vis_threshold=threshold, fancy_bbox=True
+    )
+
+    cv2.putText(
+        image,
+        f'Faces: {len(faces)}',
+        (10, 30),
+        cv2.FONT_HERSHEY_SIMPLEX,
+        1,
+        (0, 255, 0),
+        2,
+    )
+    cv2.imwrite(str(output_path), image)
+
+    return len(faces)
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Batch process images with face detection')
+    parser.add_argument('--input', type=str, required=True, help='Input directory')
+    parser.add_argument('--output', type=str, required=True, help='Output directory')
+    parser.add_argument('--detector', type=str, default='retinaface', choices=['retinaface', 'scrfd'])
+    parser.add_argument('--threshold', type=float, default=0.6, help='Visualization threshold')
+    parser.add_argument('--extensions', type=str, default='jpg,jpeg,png,bmp', help='Image extensions')
+    args = parser.parse_args()
+
+    input_path = Path(args.input)
+    output_path = Path(args.output)
+
+    if not input_path.exists():
+        print(f"Error: Input directory '{args.input}' does not exist")
+        return
+
+    output_path.mkdir(parents=True, exist_ok=True)
+
+    extensions = tuple(ext.strip() for ext in args.extensions.split(','))
+    image_files = get_image_files(input_path, extensions)
+
+    if not image_files:
+        print(f'No images found with extensions {extensions}')
+        return
+
+    print(f'Found {len(image_files)} images')
+
+    detector = RetinaFace() if args.detector == 'retinaface' else SCRFD()
+
+    success, errors, total_faces = 0, 0, 0
+
+    for img_path in tqdm(image_files, desc='Processing', unit='img'):
+        out_path = output_path / f'{img_path.stem}_detected{img_path.suffix}'
+        result = process_image(detector, img_path, out_path, args.threshold)
+
+        if result >= 0:
+            success += 1
+            total_faces += result
+        else:
+            errors += 1
+            print(f'\nFailed: {img_path.name}')
+
+    print(f'\nDone! {success} processed, {errors} errors, {total_faces} faces total')
+
+
+if __name__ == '__main__':
+    main()

scripts/download_model.py

@@ -0,0 +1,60 @@
+import argparse
+
+from uniface.constants import (
+    AgeGenderWeights,
+    ArcFaceWeights,
+    DDAMFNWeights,
+    LandmarkWeights,
+    MobileFaceWeights,
+    RetinaFaceWeights,
+    SCRFDWeights,
+    SphereFaceWeights,
+)
+from uniface.model_store import verify_model_weights
+
+MODEL_TYPES = {
+    'retinaface': RetinaFaceWeights,
+    'sphereface': SphereFaceWeights,
+    'mobileface': MobileFaceWeights,
+    'arcface': ArcFaceWeights,
+    'scrfd': SCRFDWeights,
+    'ddamfn': DDAMFNWeights,
+    'agegender': AgeGenderWeights,
+    'landmark': LandmarkWeights,
+}
+
+
+def download_models(model_enum):
+    for weight in model_enum:
+        print(f'Downloading: {weight.value}')
+        try:
+            verify_model_weights(weight)
+            print(f'  Done: {weight.value}')
+        except Exception as e:
+            print(f'  Failed: {e}')
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Download model weights')
+    parser.add_argument(
+        '--model-type',
+        type=str,
+        choices=list(MODEL_TYPES.keys()),
+        help='Model type to download. If not specified, downloads all.',
+    )
+    args = parser.parse_args()
+
+    if args.model_type:
+        print(f'Downloading {args.model_type} models...')
+        download_models(MODEL_TYPES[args.model_type])
+    else:
+        print('Downloading all models...')
+        for name, model_enum in MODEL_TYPES.items():
+            print(f'\n{name}:')
+            download_models(model_enum)
+
+    print('\nDone!')
+
+
+if __name__ == '__main__':
+    main()

scripts/release.sh

@@ -1,23 +0,0 @@
-#!/bin/bash
-
-# Exit on errors
-set -e
-
-cd "$(dirname "$0")"/..
-
-echo "Deleting existing release-related files..."
-rm -rf dist/ build/ *.egg-info
-
-pip install --upgrade pip
-pip install twine
-
-echo "Creating a package for the current release (PyPI compatible)..."
-python3 setup.py sdist bdist_wheel
-
-echo "Release package created successfully in the 'dist/' folder."
-
-
-echo "Uploading the package to PyPI..."
-twine upload dist/*
-
-echo "Release uploaded successfully!"

scripts/run_age_gender.py

@@ -0,0 +1,130 @@
+# Age and gender prediction on detected faces
+# Usage: python run_age_gender.py --image path/to/image.jpg
+#        python run_age_gender.py --webcam
+
+import argparse
+import os
+from pathlib import Path
+
+import cv2
+
+from uniface import SCRFD, AgeGender, RetinaFace
+from uniface.visualization import draw_detections
+
+
+def draw_age_gender_label(image, bbox, gender_id: int, age: int):
+    """Draw age/gender label above the bounding box."""
+    x1, y1 = int(bbox[0]), int(bbox[1])
+    gender_str = 'Female' if gender_id == 0 else 'Male'
+    text = f'{gender_str}, {age}y'
+    (tw, th), _ = cv2.getTextSize(text, cv2.FONT_HERSHEY_SIMPLEX, 0.6, 2)
+    cv2.rectangle(image, (x1, y1 - th - 10), (x1 + tw + 10, y1), (0, 255, 0), -1)
+    cv2.putText(image, text, (x1 + 5, y1 - 5), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 0), 2)
+
+
+def process_image(
+    detector,
+    age_gender,
+    image_path: str,
+    save_dir: str = 'outputs',
+    threshold: float = 0.6,
+):
+    image = cv2.imread(image_path)
+    if image is None:
+        print(f"Error: Failed to load image from '{image_path}'")
+        return
+
+    faces = detector.detect(image)
+    print(f'Detected {len(faces)} face(s)')
+
+    if not faces:
+        return
+
+    bboxes = [f['bbox'] for f in faces]
+    scores = [f['confidence'] for f in faces]
+    landmarks = [f['landmarks'] for f in faces]
+    draw_detections(
+        image=image, bboxes=bboxes, scores=scores, landmarks=landmarks, vis_threshold=threshold, fancy_bbox=True
+    )
+
+    for i, face in enumerate(faces):
+        gender_id, age = age_gender.predict(image, face['bbox'])
+        gender_str = 'Female' if gender_id == 0 else 'Male'
+        print(f'  Face {i + 1}: {gender_str}, {age} years old')
+        draw_age_gender_label(image, face['bbox'], gender_id, age)
+
+    os.makedirs(save_dir, exist_ok=True)
+    output_path = os.path.join(save_dir, f'{Path(image_path).stem}_age_gender.jpg')
+    cv2.imwrite(output_path, image)
+    print(f'Output saved: {output_path}')
+
+
+def run_webcam(detector, age_gender, threshold: float = 0.6):
+    cap = cv2.VideoCapture(0)  # 0 = default webcam
+    if not cap.isOpened():
+        print('Cannot open webcam')
+        return
+
+    print("Press 'q' to quit")
+
+    while True:
+        ret, frame = cap.read()
+        frame = cv2.flip(frame, 1)  # mirror for natural interaction
+        if not ret:
+            break
+
+        faces = detector.detect(frame)
+
+        # unpack face data for visualization
+        bboxes = [f['bbox'] for f in faces]
+        scores = [f['confidence'] for f in faces]
+        landmarks = [f['landmarks'] for f in faces]
+        draw_detections(
+            image=frame, bboxes=bboxes, scores=scores, landmarks=landmarks, vis_threshold=threshold, fancy_bbox=True
+        )
+
+        for face in faces:
+            gender_id, age = age_gender.predict(frame, face['bbox'])  # predict per face
+            draw_age_gender_label(frame, face['bbox'], gender_id, age)
+
+        cv2.putText(
+            frame,
+            f'Faces: {len(faces)}',
+            (10, 30),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            1,
+            (0, 255, 0),
+            2,
+        )
+        cv2.imshow('Age & Gender Detection', frame)
+
+        if cv2.waitKey(1) & 0xFF == ord('q'):
+            break
+
+    cap.release()
+    cv2.destroyAllWindows()
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Run age and gender detection')
+    parser.add_argument('--image', type=str, help='Path to input image')
+    parser.add_argument('--webcam', action='store_true', help='Use webcam')
+    parser.add_argument('--detector', type=str, default='retinaface', choices=['retinaface', 'scrfd'])
+    parser.add_argument('--threshold', type=float, default=0.6, help='Visualization threshold')
+    parser.add_argument('--save_dir', type=str, default='outputs')
+    args = parser.parse_args()
+
+    if not args.image and not args.webcam:
+        parser.error('Either --image or --webcam must be specified')
+
+    detector = RetinaFace() if args.detector == 'retinaface' else SCRFD()
+    age_gender = AgeGender()
+
+    if args.webcam:
+        run_webcam(detector, age_gender, args.threshold)
+    else:
+        process_image(detector, age_gender, args.image, args.save_dir, args.threshold)
+
+
+if __name__ == '__main__':
+    main()

scripts/run_detection.py

@@ -0,0 +1,110 @@
+# Face detection on image or webcam
+# Usage: python run_detection.py --image path/to/image.jpg
+#        python run_detection.py --webcam
+
+import argparse
+import os
+
+import cv2
+
+from uniface.detection import SCRFD, RetinaFace, YOLOv5Face
+from uniface.visualization import draw_detections
+
+
+def process_image(detector, image_path: str, threshold: float = 0.6, save_dir: str = 'outputs'):
+    image = cv2.imread(image_path)
+    if image is None:
+        print(f"Error: Failed to load image from '{image_path}'")
+        return
+
+    faces = detector.detect(image)
+
+    if faces:
+        bboxes = [face['bbox'] for face in faces]
+        scores = [face['confidence'] for face in faces]
+        landmarks = [face['landmarks'] for face in faces]
+        draw_detections(image, bboxes, scores, landmarks, vis_threshold=threshold)
+
+    os.makedirs(save_dir, exist_ok=True)
+    output_path = os.path.join(save_dir, f'{os.path.splitext(os.path.basename(image_path))[0]}_out.jpg')
+    cv2.imwrite(output_path, image)
+    print(f'Output saved: {output_path}')
+
+
+def run_webcam(detector, threshold: float = 0.6):
+    cap = cv2.VideoCapture(0)  # 0 = default webcam
+    if not cap.isOpened():
+        print('Cannot open webcam')
+        return
+
+    print("Press 'q' to quit")
+
+    while True:
+        ret, frame = cap.read()
+        frame = cv2.flip(frame, 1)  # mirror for natural interaction
+        if not ret:
+            break
+
+        faces = detector.detect(frame)
+
+        # unpack face data for visualization
+        bboxes = [f['bbox'] for f in faces]
+        scores = [f['confidence'] for f in faces]
+        landmarks = [f['landmarks'] for f in faces]
+        draw_detections(
+            image=frame,
+            bboxes=bboxes,
+            scores=scores,
+            landmarks=landmarks,
+            vis_threshold=threshold,
+            draw_score=True,
+            fancy_bbox=True,
+        )
+
+        cv2.putText(
+            frame,
+            f'Faces: {len(faces)}',
+            (10, 30),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            1,
+            (0, 255, 0),
+            2,
+        )
+        cv2.imshow('Face Detection', frame)
+
+        if cv2.waitKey(1) & 0xFF == ord('q'):
+            break
+
+    cap.release()
+    cv2.destroyAllWindows()
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Run face detection')
+    parser.add_argument('--image', type=str, help='Path to input image')
+    parser.add_argument('--webcam', action='store_true', help='Use webcam')
+    parser.add_argument('--method', type=str, default='retinaface', choices=['retinaface', 'scrfd', 'yolov5face'])
+    parser.add_argument('--threshold', type=float, default=0.25, help='Visualization threshold')
+    parser.add_argument('--save_dir', type=str, default='outputs')
+    args = parser.parse_args()
+
+    if not args.image and not args.webcam:
+        parser.error('Either --image or --webcam must be specified')
+
+    if args.method == 'retinaface':
+        detector = RetinaFace()
+    elif args.method == 'scrfd':
+        detector = SCRFD()
+    else:
+        from uniface.constants import YOLOv5FaceWeights
+
+        detector = YOLOv5Face(model_name=YOLOv5FaceWeights.YOLOV5M)
+
+    if args.webcam:
+        run_webcam(detector, args.threshold)
+    else:
+        process_image(detector, args.image, args.threshold, args.save_dir)
+
+
+if __name__ == '__main__':
+    main()

scripts/run_emotion.py

@@ -0,0 +1,126 @@
+# Emotion detection on detected faces
+# Usage: python run_emotion.py --image path/to/image.jpg
+#        python run_emotion.py --webcam
+
+import argparse
+import os
+from pathlib import Path
+
+import cv2
+
+from uniface import SCRFD, Emotion, RetinaFace
+from uniface.visualization import draw_detections
+
+
+def draw_emotion_label(image, bbox, emotion: str, confidence: float):
+    """Draw emotion label above the bounding box."""
+    x1, y1 = int(bbox[0]), int(bbox[1])
+    text = f'{emotion} ({confidence:.2f})'
+    (tw, th), _ = cv2.getTextSize(text, cv2.FONT_HERSHEY_SIMPLEX, 0.6, 2)
+    cv2.rectangle(image, (x1, y1 - th - 10), (x1 + tw + 10, y1), (255, 0, 0), -1)
+    cv2.putText(image, text, (x1 + 5, y1 - 5), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (255, 255, 255), 2)
+
+
+def process_image(
+    detector,
+    emotion_predictor,
+    image_path: str,
+    save_dir: str = 'outputs',
+    threshold: float = 0.6,
+):
+    image = cv2.imread(image_path)
+    if image is None:
+        print(f"Error: Failed to load image from '{image_path}'")
+        return
+
+    faces = detector.detect(image)
+    print(f'Detected {len(faces)} face(s)')
+
+    if not faces:
+        return
+
+    bboxes = [f['bbox'] for f in faces]
+    scores = [f['confidence'] for f in faces]
+    landmarks = [f['landmarks'] for f in faces]
+    draw_detections(
+        image=image, bboxes=bboxes, scores=scores, landmarks=landmarks, vis_threshold=threshold, fancy_bbox=True
+    )
+
+    for i, face in enumerate(faces):
+        emotion, confidence = emotion_predictor.predict(image, face['landmarks'])
+        print(f'  Face {i + 1}: {emotion} (confidence: {confidence:.3f})')
+        draw_emotion_label(image, face['bbox'], emotion, confidence)
+
+    os.makedirs(save_dir, exist_ok=True)
+    output_path = os.path.join(save_dir, f'{Path(image_path).stem}_emotion.jpg')
+    cv2.imwrite(output_path, image)
+    print(f'Output saved: {output_path}')
+
+
+def run_webcam(detector, emotion_predictor, threshold: float = 0.6):
+    cap = cv2.VideoCapture(0)  # 0 = default webcam
+    if not cap.isOpened():
+        print('Cannot open webcam')
+        return
+
+    print("Press 'q' to quit")
+
+    while True:
+        ret, frame = cap.read()
+        frame = cv2.flip(frame, 1)  # mirror for natural interaction
+        if not ret:
+            break
+
+        faces = detector.detect(frame)
+
+        # unpack face data for visualization
+        bboxes = [f['bbox'] for f in faces]
+        scores = [f['confidence'] for f in faces]
+        landmarks = [f['landmarks'] for f in faces]
+        draw_detections(frame, bboxes, scores, landmarks, vis_threshold=threshold)
+
+        for face in faces:
+            emotion, confidence = emotion_predictor.predict(frame, face['landmarks'])
+            draw_emotion_label(frame, face['bbox'], emotion, confidence)
+
+        cv2.putText(
+            frame,
+            f'Faces: {len(faces)}',
+            (10, 30),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            1,
+            (0, 255, 0),
+            2,
+        )
+        cv2.imshow('Emotion Detection', frame)
+
+        if cv2.waitKey(1) & 0xFF == ord('q'):
+            break
+
+    cap.release()
+    cv2.destroyAllWindows()
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Run emotion detection')
+    parser.add_argument('--image', type=str, help='Path to input image')
+    parser.add_argument('--webcam', action='store_true', help='Use webcam')
+    parser.add_argument('--detector', type=str, default='retinaface', choices=['retinaface', 'scrfd'])
+    parser.add_argument('--threshold', type=float, default=0.6, help='Visualization threshold')
+    parser.add_argument('--save_dir', type=str, default='outputs')
+    args = parser.parse_args()
+
+    if not args.image and not args.webcam:
+        parser.error('Either --image or --webcam must be specified')
+
+    detector = RetinaFace() if args.detector == 'retinaface' else SCRFD()
+    emotion_predictor = Emotion()
+
+    if args.webcam:
+        run_webcam(detector, emotion_predictor, args.threshold)
+    else:
+        process_image(detector, emotion_predictor, args.image, args.save_dir, args.threshold)
+
+
+if __name__ == '__main__':
+    main()

scripts/run_face_analyzer.py

@@ -0,0 +1,116 @@
+# Face analysis using FaceAnalyzer
+# Usage: python run_face_analyzer.py --image path/to/image.jpg
+
+import argparse
+import os
+from pathlib import Path
+
+import cv2
+import numpy as np
+
+from uniface import AgeGender, ArcFace, FaceAnalyzer, RetinaFace
+from uniface.visualization import draw_detections
+
+
+def draw_face_info(image, face, face_id):
+    """Draw face ID and attributes above bounding box."""
+    x1, y1, x2, y2 = map(int, face.bbox)
+    lines = [f'ID: {face_id}', f'Conf: {face.confidence:.2f}']
+    if face.age and face.sex:
+        lines.append(f'{face.sex}, {face.age}y')
+
+    for i, line in enumerate(lines):
+        y_pos = y1 - 10 - (len(lines) - 1 - i) * 25
+        if y_pos < 20:
+            y_pos = y2 + 20 + i * 25
+        (tw, th), _ = cv2.getTextSize(line, cv2.FONT_HERSHEY_SIMPLEX, 0.6, 2)
+        cv2.rectangle(image, (x1, y_pos - th - 5), (x1 + tw + 10, y_pos + 5), (0, 255, 0), -1)
+        cv2.putText(image, line, (x1 + 5, y_pos), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 0), 2)
+
+
+def process_image(analyzer, image_path: str, save_dir: str = 'outputs', show_similarity: bool = True):
+    image = cv2.imread(image_path)
+    if image is None:
+        print(f"Error: Failed to load image from '{image_path}'")
+        return
+
+    faces = analyzer.analyze(image)
+    print(f'Detected {len(faces)} face(s)')
+
+    if not faces:
+        return
+
+    for i, face in enumerate(faces, 1):
+        info = f'  Face {i}: {face.sex}, {face.age}y' if face.age and face.sex else f'  Face {i}'
+        if face.embedding is not None:
+            info += f' (embedding: {face.embedding.shape})'
+        print(info)
+
+    if show_similarity and len(faces) >= 2:
+        print('\nSimilarity Matrix:')
+        n = len(faces)
+        sim_matrix = np.zeros((n, n))
+
+        for i in range(n):
+            for j in range(i, n):
+                if i == j:
+                    sim_matrix[i][j] = 1.0
+                else:
+                    sim = faces[i].compute_similarity(faces[j])
+                    sim_matrix[i][j] = sim
+                    sim_matrix[j][i] = sim
+
+        print('     ', end='')
+        for i in range(n):
+            print(f'  F{i + 1:2d}  ', end='')
+        print('\n     ' + '-' * (7 * n))
+
+        for i in range(n):
+            print(f'F{i + 1:2d} | ', end='')
+            for j in range(n):
+                print(f'{sim_matrix[i][j]:6.3f} ', end='')
+            print()
+
+        pairs = [(i, j, sim_matrix[i][j]) for i in range(n) for j in range(i + 1, n)]
+        pairs.sort(key=lambda x: x[2], reverse=True)
+
+        print('\nTop matches (>0.4 = same person):')
+        for i, j, sim in pairs[:3]:
+            status = 'Same' if sim > 0.4 else 'Different'
+            print(f'  Face {i + 1} ↔ Face {j + 1}: {sim:.3f} ({status})')
+
+    bboxes = [f.bbox for f in faces]
+    scores = [f.confidence for f in faces]
+    landmarks = [f.landmarks for f in faces]
+    draw_detections(image=image, bboxes=bboxes, scores=scores, landmarks=landmarks, fancy_bbox=True)
+
+    for i, face in enumerate(faces, 1):
+        draw_face_info(image, face, i)
+
+    os.makedirs(save_dir, exist_ok=True)
+    output_path = os.path.join(save_dir, f'{Path(image_path).stem}_analysis.jpg')
+    cv2.imwrite(output_path, image)
+    print(f'Output saved: {output_path}')
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Face analysis with detection, recognition, and attributes')
+    parser.add_argument('--image', type=str, required=True, help='Path to input image')
+    parser.add_argument('--save_dir', type=str, default='outputs', help='Output directory')
+    parser.add_argument('--no-similarity', action='store_true', help='Skip similarity matrix computation')
+    args = parser.parse_args()
+
+    if not os.path.exists(args.image):
+        print(f'Error: Image not found: {args.image}')
+        return
+
+    detector = RetinaFace()
+    recognizer = ArcFace()
+    age_gender = AgeGender()
+    analyzer = FaceAnalyzer(detector, recognizer, age_gender)
+
+    process_image(analyzer, args.image, args.save_dir, show_similarity=not args.no_similarity)
+
+
+if __name__ == '__main__':
+    main()

scripts/run_face_parsing.py

@@ -0,0 +1,126 @@
+# Face parsing on detected faces
+# Usage: python run_face_parsing.py --image path/to/image.jpg
+#        python run_face_parsing.py --webcam
+
+import argparse
+import os
+from pathlib import Path
+
+import cv2
+
+from uniface import RetinaFace
+from uniface.constants import ParsingWeights
+from uniface.parsing import BiSeNet
+from uniface.visualization import vis_parsing_maps
+
+
+def process_image(detector, parser, image_path: str, save_dir: str = 'outputs'):
+    image = cv2.imread(image_path)
+    if image is None:
+        print(f"Error: Failed to load image from '{image_path}'")
+        return
+
+    faces = detector.detect(image)
+    print(f'Detected {len(faces)} face(s)')
+
+    result_image = image.copy()
+
+    for i, face in enumerate(faces):
+        bbox = face['bbox']
+        x1, y1, x2, y2 = map(int, bbox[:4])
+        face_crop = image[y1:y2, x1:x2]
+
+        if face_crop.size == 0:
+            continue
+
+        # Parse the face
+        mask = parser.parse(face_crop)
+        print(f'  Face {i + 1}: parsed with {len(set(mask.flatten()))} unique classes')
+
+        # Visualize the parsing result
+        face_crop_rgb = cv2.cvtColor(face_crop, cv2.COLOR_BGR2RGB)
+        vis_result = vis_parsing_maps(face_crop_rgb, mask, save_image=False)
+
+        # Place the visualization back on the original image
+        result_image[y1:y2, x1:x2] = vis_result
+
+        # Draw bounding box
+        cv2.rectangle(result_image, (x1, y1), (x2, y2), (0, 255, 0), 2)
+
+    os.makedirs(save_dir, exist_ok=True)
+    output_path = os.path.join(save_dir, f'{Path(image_path).stem}_parsing.jpg')
+    cv2.imwrite(output_path, result_image)
+    print(f'Output saved: {output_path}')
+
+
+def run_webcam(detector, parser):
+    cap = cv2.VideoCapture(0)
+    if not cap.isOpened():
+        print('Cannot open webcam')
+        return
+
+    print("Press 'q' to quit")
+
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            break
+
+        frame = cv2.flip(frame, 1)
+        faces = detector.detect(frame)
+
+        for face in faces:
+            bbox = face['bbox']
+            x1, y1, x2, y2 = map(int, bbox[:4])
+            face_crop = frame[y1:y2, x1:x2]
+
+            if face_crop.size == 0:
+                continue
+
+            # Parse the face
+            mask = parser.parse(face_crop)
+
+            # Visualize the parsing result
+            face_crop_rgb = cv2.cvtColor(face_crop, cv2.COLOR_BGR2RGB)
+            vis_result = vis_parsing_maps(face_crop_rgb, mask, save_image=False)
+
+            # Place the visualization back on the frame
+            frame[y1:y2, x1:x2] = vis_result
+
+            # Draw bounding box
+            cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
+
+        cv2.putText(frame, f'Faces: {len(faces)}', (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)
+        cv2.imshow('Face Parsing', frame)
+
+        if cv2.waitKey(1) & 0xFF == ord('q'):
+            break
+
+    cap.release()
+    cv2.destroyAllWindows()
+
+
+def main():
+    parser_arg = argparse.ArgumentParser(description='Run face parsing')
+    parser_arg.add_argument('--image', type=str, help='Path to input image')
+    parser_arg.add_argument('--webcam', action='store_true', help='Use webcam')
+    parser_arg.add_argument('--save_dir', type=str, default='outputs')
+    parser_arg.add_argument(
+        '--model', type=str, default=ParsingWeights.RESNET18, choices=[ParsingWeights.RESNET18, ParsingWeights.RESNET34]
+    )
+    args = parser_arg.parse_args()
+
+    if not args.image and not args.webcam:
+        parser_arg.error('Either --image or --webcam must be specified')
+
+    detector = RetinaFace()
+    parser = BiSeNet(model_name=ParsingWeights.RESNET34)
+
+    if args.webcam:
+        run_webcam(detector, parser)
+    else:
+        process_image(detector, parser, args.image, args.save_dir)
+
+
+if __name__ == '__main__':
+    main()

scripts/run_face_search.py

@@ -0,0 +1,97 @@
+# Real-time face search: match webcam faces against a reference image
+# Usage: python run_face_search.py --image reference.jpg
+
+import argparse
+
+import cv2
+import numpy as np
+
+from uniface.detection import SCRFD, RetinaFace
+from uniface.face_utils import compute_similarity
+from uniface.recognition import ArcFace, MobileFace, SphereFace
+
+
+def get_recognizer(name: str):
+    if name == 'arcface':
+        return ArcFace()
+    elif name == 'mobileface':
+        return MobileFace()
+    else:
+        return SphereFace()
+
+
+def extract_reference_embedding(detector, recognizer, image_path: str) -> np.ndarray:
+    image = cv2.imread(image_path)
+    if image is None:
+        raise RuntimeError(f'Failed to load image: {image_path}')
+
+    faces = detector.detect(image)
+    if not faces:
+        raise RuntimeError('No faces found in reference image.')
+
+    landmarks = faces[0]['landmarks']
+    return recognizer.get_normalized_embedding(image, landmarks)
+
+
+def run_webcam(detector, recognizer, ref_embedding: np.ndarray, threshold: float = 0.4):
+    cap = cv2.VideoCapture(0)  # 0 = default webcam
+    if not cap.isOpened():
+        raise RuntimeError('Webcam could not be opened.')
+
+    print("Press 'q' to quit")
+
+    while True:
+        ret, frame = cap.read()
+        frame = cv2.flip(frame, 1)  # mirror for natural interaction
+        if not ret:
+            break
+
+        faces = detector.detect(frame)
+
+        for face in faces:
+            bbox = face['bbox']
+            landmarks = face['landmarks']
+            x1, y1, x2, y2 = map(int, bbox)
+
+            embedding = recognizer.get_normalized_embedding(frame, landmarks)
+            sim = compute_similarity(ref_embedding, embedding)  # compare with reference
+
+            # green = match, red = unknown
+            label = f'Match ({sim:.2f})' if sim > threshold else f'Unknown ({sim:.2f})'
+            color = (0, 255, 0) if sim > threshold else (0, 0, 255)
+
+            cv2.rectangle(frame, (x1, y1), (x2, y2), color, 2)
+            cv2.putText(frame, label, (x1, y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.6, color, 2)
+
+        cv2.imshow('Face Recognition', frame)
+        if cv2.waitKey(1) & 0xFF == ord('q'):
+            break
+
+    cap.release()
+    cv2.destroyAllWindows()
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Face search using a reference image')
+    parser.add_argument('--image', type=str, required=True, help='Reference face image')
+    parser.add_argument('--threshold', type=float, default=0.4, help='Match threshold')
+    parser.add_argument('--detector', type=str, default='scrfd', choices=['retinaface', 'scrfd'])
+    parser.add_argument(
+        '--recognizer',
+        type=str,
+        default='arcface',
+        choices=['arcface', 'mobileface', 'sphereface'],
+    )
+    args = parser.parse_args()
+
+    detector = RetinaFace() if args.detector == 'retinaface' else SCRFD()
+    recognizer = get_recognizer(args.recognizer)
+
+    print(f'Loading reference: {args.image}')
+    ref_embedding = extract_reference_embedding(detector, recognizer, args.image)
+
+    run_webcam(detector, recognizer, ref_embedding, args.threshold)
+
+
+if __name__ == '__main__':
+    main()

scripts/run_gaze_estimation.py

@@ -0,0 +1,104 @@
+# Gaze estimation on detected faces
+# Usage: python run_gaze_estimation.py --image path/to/image.jpg
+#        python run_gaze_estimation.py --webcam
+
+import argparse
+import os
+from pathlib import Path
+
+import cv2
+import numpy as np
+
+from uniface import RetinaFace
+from uniface.gaze import MobileGaze
+from uniface.visualization import draw_gaze
+
+
+def process_image(detector, gaze_estimator, image_path: str, save_dir: str = 'outputs'):
+    image = cv2.imread(image_path)
+    if image is None:
+        print(f"Error: Failed to load image from '{image_path}'")
+        return
+
+    faces = detector.detect(image)
+    print(f'Detected {len(faces)} face(s)')
+
+    for i, face in enumerate(faces):
+        bbox = face['bbox']
+        x1, y1, x2, y2 = map(int, bbox[:4])
+        face_crop = image[y1:y2, x1:x2]
+
+        if face_crop.size == 0:
+            continue
+
+        pitch, yaw = gaze_estimator.estimate(face_crop)
+        print(f'  Face {i + 1}: pitch={np.degrees(pitch):.1f}°, yaw={np.degrees(yaw):.1f}°')
+
+        # Draw both bbox and gaze arrow with angle text
+        draw_gaze(image, bbox, pitch, yaw, draw_angles=True)
+
+    os.makedirs(save_dir, exist_ok=True)
+    output_path = os.path.join(save_dir, f'{Path(image_path).stem}_gaze.jpg')
+    cv2.imwrite(output_path, image)
+    print(f'Output saved: {output_path}')
+
+
+def run_webcam(detector, gaze_estimator):
+    cap = cv2.VideoCapture(0)
+    if not cap.isOpened():
+        print('Cannot open webcam')
+        return
+
+    print("Press 'q' to quit")
+
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            break
+
+        frame = cv2.flip(frame, 1)
+        faces = detector.detect(frame)
+
+        for face in faces:
+            bbox = face['bbox']
+            x1, y1, x2, y2 = map(int, bbox[:4])
+            face_crop = frame[y1:y2, x1:x2]
+
+            if face_crop.size == 0:
+                continue
+
+            pitch, yaw = gaze_estimator.estimate(face_crop)
+            # Draw both bbox and gaze arrow
+            draw_gaze(frame, bbox, pitch, yaw)
+
+        cv2.putText(frame, f'Faces: {len(faces)}', (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)
+        cv2.imshow('Gaze Estimation', frame)
+
+        if cv2.waitKey(1) & 0xFF == ord('q'):
+            break
+
+    cap.release()
+    cv2.destroyAllWindows()
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Run gaze estimation')
+    parser.add_argument('--image', type=str, help='Path to input image')
+    parser.add_argument('--webcam', action='store_true', help='Use webcam')
+    parser.add_argument('--save_dir', type=str, default='outputs')
+    args = parser.parse_args()
+
+    if not args.image and not args.webcam:
+        parser.error('Either --image or --webcam must be specified')
+
+    detector = RetinaFace()
+    gaze_estimator = MobileGaze()
+
+    if args.webcam:
+        run_webcam(detector, gaze_estimator)
+    else:
+        process_image(detector, gaze_estimator, args.image, args.save_dir)
+
+
+if __name__ == '__main__':
+    main()

scripts/run_landmarks.py

@@ -0,0 +1,117 @@
+# 106-point facial landmark detection
+# Usage: python run_landmarks.py --image path/to/image.jpg
+#        python run_landmarks.py --webcam
+
+import argparse
+import os
+from pathlib import Path
+
+import cv2
+
+from uniface import SCRFD, Landmark106, RetinaFace
+
+
+def process_image(detector, landmarker, image_path: str, save_dir: str = 'outputs'):
+    image = cv2.imread(image_path)
+    if image is None:
+        print(f"Error: Failed to load image from '{image_path}'")
+        return
+
+    faces = detector.detect(image)
+    print(f'Detected {len(faces)} face(s)')
+
+    if not faces:
+        return
+
+    for i, face in enumerate(faces):
+        bbox = face['bbox']
+        x1, y1, x2, y2 = map(int, bbox)
+        cv2.rectangle(image, (x1, y1), (x2, y2), (0, 255, 0), 2)
+
+        landmarks = landmarker.get_landmarks(image, bbox)
+        print(f'  Face {i + 1}: {len(landmarks)} landmarks')
+
+        for x, y in landmarks.astype(int):
+            cv2.circle(image, (x, y), 1, (0, 255, 0), -1)
+
+        cv2.putText(
+            image,
+            f'Face {i + 1}',
+            (x1, y1 - 10),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.5,
+            (0, 255, 0),
+            2,
+        )
+
+    os.makedirs(save_dir, exist_ok=True)
+    output_path = os.path.join(save_dir, f'{Path(image_path).stem}_landmarks.jpg')
+    cv2.imwrite(output_path, image)
+    print(f'Output saved: {output_path}')
+
+
+def run_webcam(detector, landmarker):
+    cap = cv2.VideoCapture(0)  # 0 = default webcam
+    if not cap.isOpened():
+        print('Cannot open webcam')
+        return
+
+    print("Press 'q' to quit")
+
+    while True:
+        ret, frame = cap.read()
+        frame = cv2.flip(frame, 1)  # mirror for natural interaction
+        if not ret:
+            break
+
+        faces = detector.detect(frame)
+
+        for face in faces:
+            bbox = face['bbox']
+            x1, y1, x2, y2 = map(int, bbox)
+            cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
+
+            landmarks = landmarker.get_landmarks(frame, bbox)  # 106 points
+            for x, y in landmarks.astype(int):
+                cv2.circle(frame, (x, y), 1, (0, 255, 0), -1)
+
+        cv2.putText(
+            frame,
+            f'Faces: {len(faces)}',
+            (10, 30),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            1,
+            (0, 255, 0),
+            2,
+        )
+        cv2.imshow('106-Point Landmarks', frame)
+
+        if cv2.waitKey(1) & 0xFF == ord('q'):
+            break
+
+    cap.release()
+    cv2.destroyAllWindows()
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Run facial landmark detection')
+    parser.add_argument('--image', type=str, help='Path to input image')
+    parser.add_argument('--webcam', action='store_true', help='Use webcam')
+    parser.add_argument('--detector', type=str, default='retinaface', choices=['retinaface', 'scrfd'])
+    parser.add_argument('--save_dir', type=str, default='outputs')
+    args = parser.parse_args()
+
+    if not args.image and not args.webcam:
+        parser.error('Either --image or --webcam must be specified')
+
+    detector = RetinaFace() if args.detector == 'retinaface' else SCRFD()
+    landmarker = Landmark106()
+
+    if args.webcam:
+        run_webcam(detector, landmarker)
+    else:
+        process_image(detector, landmarker, args.image, args.save_dir)
+
+
+if __name__ == '__main__':
+    main()

scripts/run_recognition.py

@@ -0,0 +1,103 @@
+# Face recognition: extract embeddings or compare two faces
+# Usage: python run_recognition.py --image path/to/image.jpg
+#        python run_recognition.py --image1 face1.jpg --image2 face2.jpg
+
+import argparse
+
+import cv2
+import numpy as np
+
+from uniface.detection import SCRFD, RetinaFace
+from uniface.face_utils import compute_similarity
+from uniface.recognition import ArcFace, MobileFace, SphereFace
+
+
+def get_recognizer(name: str):
+    if name == 'arcface':
+        return ArcFace()
+    elif name == 'mobileface':
+        return MobileFace()
+    else:
+        return SphereFace()
+
+
+def run_inference(detector, recognizer, image_path: str):
+    image = cv2.imread(image_path)
+    if image is None:
+        print(f"Error: Failed to load image from '{image_path}'")
+        return
+
+    faces = detector.detect(image)
+    if not faces:
+        print('No faces detected.')
+        return
+
+    print(f'Detected {len(faces)} face(s). Extracting embedding for the first face...')
+
+    landmarks = faces[0]['landmarks']  # 5-point landmarks for alignment (already np.ndarray)
+    embedding = recognizer.get_embedding(image, landmarks)
+    norm_embedding = recognizer.get_normalized_embedding(image, landmarks)  # L2 normalized
+
+    print(f'  Embedding shape: {embedding.shape}')
+    print(f'  L2 norm (raw): {np.linalg.norm(embedding):.4f}')
+    print(f'  L2 norm (normalized): {np.linalg.norm(norm_embedding):.4f}')
+
+
+def compare_faces(detector, recognizer, image1_path: str, image2_path: str, threshold: float = 0.35):
+    img1 = cv2.imread(image1_path)
+    img2 = cv2.imread(image2_path)
+
+    if img1 is None or img2 is None:
+        print('Error: Failed to load one or both images')
+        return
+
+    faces1 = detector.detect(img1)
+    faces2 = detector.detect(img2)
+
+    if not faces1 or not faces2:
+        print('Error: No faces detected in one or both images')
+        return
+
+    landmarks1 = faces1[0]['landmarks']
+    landmarks2 = faces2[0]['landmarks']
+
+    embedding1 = recognizer.get_normalized_embedding(img1, landmarks1)
+    embedding2 = recognizer.get_normalized_embedding(img2, landmarks2)
+
+    # cosine similarity for normalized embeddings
+    similarity = compute_similarity(embedding1, embedding2, normalized=True)
+    is_match = similarity > threshold
+
+    print(f'Similarity: {similarity:.4f}')
+    print(f'Result: {"Same person" if is_match else "Different person"} (threshold: {threshold})')
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Face recognition and comparison')
+    parser.add_argument('--image', type=str, help='Single image for embedding extraction')
+    parser.add_argument('--image1', type=str, help='First image for comparison')
+    parser.add_argument('--image2', type=str, help='Second image for comparison')
+    parser.add_argument('--threshold', type=float, default=0.35, help='Similarity threshold')
+    parser.add_argument('--detector', type=str, default='retinaface', choices=['retinaface', 'scrfd'])
+    parser.add_argument(
+        '--recognizer',
+        type=str,
+        default='arcface',
+        choices=['arcface', 'mobileface', 'sphereface'],
+    )
+    args = parser.parse_args()
+
+    detector = RetinaFace() if args.detector == 'retinaface' else SCRFD()
+    recognizer = get_recognizer(args.recognizer)
+
+    if args.image1 and args.image2:
+        compare_faces(detector, recognizer, args.image1, args.image2, args.threshold)
+    elif args.image:
+        run_inference(detector, recognizer, args.image)
+    else:
+        print('Error: Provide --image or both --image1 and --image2')
+        parser.print_help()
+
+
+if __name__ == '__main__':
+    main()

scripts/run_video_detection.py

@@ -0,0 +1,109 @@
+# Face detection on video files
+# Usage: python run_video_detection.py --input video.mp4 --output output.mp4
+
+import argparse
+from pathlib import Path
+
+import cv2
+from tqdm import tqdm
+
+from uniface import SCRFD, RetinaFace
+from uniface.visualization import draw_detections
+
+
+def process_video(
+    detector,
+    input_path: str,
+    output_path: str,
+    threshold: float = 0.6,
+    show_preview: bool = False,
+):
+    cap = cv2.VideoCapture(input_path)
+    if not cap.isOpened():
+        print(f"Error: Cannot open video file '{input_path}'")
+        return
+
+    # get video properties
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    fps = cap.get(cv2.CAP_PROP_FPS)
+    width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+
+    print(f'Input: {input_path} ({width}x{height}, {fps:.1f} fps, {total_frames} frames)')
+    print(f'Output: {output_path}')
+
+    fourcc = cv2.VideoWriter_fourcc(*'mp4v')  # codec for .mp4
+    out = cv2.VideoWriter(output_path, fourcc, fps, (width, height))
+
+    if not out.isOpened():
+        print(f"Error: Cannot create output video '{output_path}'")
+        cap.release()
+        return
+
+    frame_count = 0
+    total_faces = 0
+
+    for _ in tqdm(range(total_frames), desc='Processing', unit='frames'):
+        ret, frame = cap.read()
+        if not ret:
+            break
+
+        frame_count += 1
+        faces = detector.detect(frame)
+        total_faces += len(faces)
+
+        bboxes = [f['bbox'] for f in faces]
+        scores = [f['confidence'] for f in faces]
+        landmarks = [f['landmarks'] for f in faces]
+        draw_detections(
+            image=frame, bboxes=bboxes, scores=scores, landmarks=landmarks, vis_threshold=threshold, fancy_bbox=True
+        )
+
+        cv2.putText(
+            frame,
+            f'Faces: {len(faces)}',
+            (10, 30),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            1,
+            (0, 255, 0),
+            2,
+        )
+        out.write(frame)
+
+        if show_preview:
+            cv2.imshow("Processing - Press 'q' to cancel", frame)
+            if cv2.waitKey(1) & 0xFF == ord('q'):
+                print('\nCancelled by user')
+                break
+
+    cap.release()
+    out.release()
+    if show_preview:
+        cv2.destroyAllWindows()
+
+    avg_faces = total_faces / frame_count if frame_count > 0 else 0
+    print(f'\nDone! {frame_count} frames, {total_faces} faces ({avg_faces:.1f} avg/frame)')
+    print(f'Saved: {output_path}')
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Process video with face detection')
+    parser.add_argument('--input', type=str, required=True, help='Input video path')
+    parser.add_argument('--output', type=str, required=True, help='Output video path')
+    parser.add_argument('--detector', type=str, default='retinaface', choices=['retinaface', 'scrfd'])
+    parser.add_argument('--threshold', type=float, default=0.6, help='Visualization threshold')
+    parser.add_argument('--preview', action='store_true', help='Show live preview')
+    args = parser.parse_args()
+
+    if not Path(args.input).exists():
+        print(f"Error: Input file '{args.input}' does not exist")
+        return
+
+    Path(args.output).parent.mkdir(parents=True, exist_ok=True)
+
+    detector = RetinaFace() if args.detector == 'retinaface' else SCRFD()
+    process_video(detector, args.input, args.output, args.threshold, args.preview)
+
+
+if __name__ == '__main__':
+    main()

scripts/sha256_generate.py

@@ -0,0 +1,28 @@
+import argparse
+import hashlib
+from pathlib import Path
+
+
+def compute_sha256(file_path: Path, chunk_size: int = 8192) -> str:
+    sha256_hash = hashlib.sha256()
+    with file_path.open('rb') as f:
+        for chunk in iter(lambda: f.read(chunk_size), b''):
+            sha256_hash.update(chunk)
+    return sha256_hash.hexdigest()
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Compute SHA256 hash of a file')
+    parser.add_argument('file', type=Path, help='Path to file')
+    args = parser.parse_args()
+
+    if not args.file.exists() or not args.file.is_file():
+        print(f'File does not exist: {args.file}')
+        return
+
+    sha256 = compute_sha256(args.file)
+    print(f"SHA256 hash for '{args.file.name}':\n{sha256}")
+
+
+if __name__ == '__main__':
+    main()

setup.py

@@ -1,43 +0,0 @@
-import os
-from setuptools import setup, find_packages
-
-# Read the README file for the long description
-long_description = ""
-if os.path.exists("README.md"):
-    with open("README.md", "r", encoding="utf-8") as f:
-        long_description = f.read()
-
-setup(
-    name="uniface",
-    version="0.1.1",
-    packages=find_packages(),
-    install_requires=[
-        "numpy",
-        "opencv-python",
-        "onnx",
-        "onnxruntime",
-        "requests",
-        "torch"
-    ],
-    extras_require={
-        "dev": ["pytest"],
-    },
-    description="UniFace: A Comprehensive Library for Face Detection, Recognition, Landmark Analysis, Age, and Gender Detection",
-    long_description=long_description,
-    long_description_content_type="text/markdown",
-    author="Yakhyokhuja Valikhujaev",
-    author_email="yakhyo9696@gmail.com",
-    url="https://github.com/yakhyo/uniface",
-    license="MIT",
-    classifiers=[
-        "Programming Language :: Python :: 3",
-        "Programming Language :: Python :: 3.8",
-        "Programming Language :: Python :: 3.9",
-        "Programming Language :: Python :: 3.10",
-        "License :: OSI Approved :: MIT License",
-        "Operating System :: OS Independent",
-        "Topic :: Software Development :: Libraries :: Python Modules",
-    ],
-    keywords="face detection, face recognition, facial landmark, facial attribute, onnx, opencv, retinaface",
-    python_requires=">=3.8",
-)

test.py

@@ -1,57 +0,0 @@
-import os
-import cv2
-import numpy as np
-
-from uniface import RetinaFace, draw_detections
-
-
-def run_inference(image_path, save_image=False, vis_threshold=0.6):
-    """
-    Perform inference on an image, draw detections, and optionally save the output image.
-
-    Args:
-        image_path (str): Path to the input image.
-        save_image (bool): Whether to save the output image with detections.
-        vis_threshold (float): Confidence threshold for displaying detections.
-    """
-    # Load the image
-    original_image = cv2.imread(image_path)
-    if original_image is None:
-        print(f"Error: Could not read image from {image_path}")
-        return
-
-    # Perform face detection
-    boxes, landmarks = retinaface_inference.detect(original_image)
-
-    # Draw detections on the image
-    draw_detections(original_image, (boxes, landmarks), vis_threshold)
-
-    # Save the output image if requested
-    if save_image:
-        im_name = os.path.splitext(os.path.basename(image_path))[0]
-        save_name = f"{im_name}_out.jpg"
-        cv2.imwrite(save_name, original_image)
-        print(f"Image saved at '{save_name}'")
-
-
-if __name__ == '__main__':
-    import time
-
-    # Initialize and run the ONNX inference
-    retinaface_inference = RetinaFace(
-        model="retinaface_mnet_v2",
-        conf_thresh=0.5,
-        pre_nms_topk=5000,
-        nms_thresh=0.4,
-        post_nms_topk=750,
-    )
-
-    img_path = "assets/test.jpg"
-    avg = 0
-    for _ in range(50):
-        st = time.time()
-        run_inference(img_path, save_image=True, vis_threshold=0.6)
-        d = time.time() - st
-        print(d)
-        avg += d
-    print("avg", avg / 50)

tests/test_age_gender.py

@@ -0,0 +1,117 @@
+import numpy as np
+import pytest
+
+from uniface.attribute import AgeGender
+
+
+@pytest.fixture
+def age_gender_model():
+    return AgeGender()
+
+
+@pytest.fixture
+def mock_image():
+    return np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+
+
+@pytest.fixture
+def mock_bbox():
+    return [100, 100, 300, 300]
+
+
+def test_model_initialization(age_gender_model):
+    assert age_gender_model is not None, 'AgeGender model initialization failed.'
+
+
+def test_prediction_output_format(age_gender_model, mock_image, mock_bbox):
+    gender_id, age = age_gender_model.predict(mock_image, mock_bbox)
+    assert isinstance(gender_id, int), f'Gender ID should be int, got {type(gender_id)}'
+    assert isinstance(age, int), f'Age should be int, got {type(age)}'
+
+
+def test_gender_values(age_gender_model, mock_image, mock_bbox):
+    gender_id, age = age_gender_model.predict(mock_image, mock_bbox)
+    assert gender_id in [0, 1], f'Gender ID should be 0 (Female) or 1 (Male), got {gender_id}'
+
+
+def test_age_range(age_gender_model, mock_image, mock_bbox):
+    gender_id, age = age_gender_model.predict(mock_image, mock_bbox)
+    assert 0 <= age <= 120, f'Age should be between 0 and 120, got {age}'
+
+
+def test_different_bbox_sizes(age_gender_model, mock_image):
+    test_bboxes = [
+        [50, 50, 150, 150],
+        [100, 100, 300, 300],
+        [50, 50, 400, 400],
+    ]
+
+    for bbox in test_bboxes:
+        gender_id, age = age_gender_model.predict(mock_image, bbox)
+        assert gender_id in [0, 1], f'Failed for bbox {bbox}'
+        assert 0 <= age <= 120, f'Age out of range for bbox {bbox}'
+
+
+def test_different_image_sizes(age_gender_model, mock_bbox):
+    test_sizes = [(480, 640, 3), (720, 1280, 3), (1080, 1920, 3)]
+
+    for size in test_sizes:
+        mock_image = np.random.randint(0, 255, size, dtype=np.uint8)
+        gender_id, age = age_gender_model.predict(mock_image, mock_bbox)
+        assert gender_id in [0, 1], f'Failed for image size {size}'
+        assert 0 <= age <= 120, f'Age out of range for image size {size}'
+
+
+def test_consistency(age_gender_model, mock_image, mock_bbox):
+    gender_id1, age1 = age_gender_model.predict(mock_image, mock_bbox)
+    gender_id2, age2 = age_gender_model.predict(mock_image, mock_bbox)
+
+    assert gender_id1 == gender_id2, 'Same input should produce same gender prediction'
+    assert age1 == age2, 'Same input should produce same age prediction'
+
+
+def test_bbox_list_format(age_gender_model, mock_image):
+    bbox_list = [100, 100, 300, 300]
+    gender_id, age = age_gender_model.predict(mock_image, bbox_list)
+    assert gender_id in [0, 1], 'Should work with bbox as list'
+    assert 0 <= age <= 120, 'Age should be in valid range'
+
+
+def test_bbox_array_format(age_gender_model, mock_image):
+    bbox_array = np.array([100, 100, 300, 300])
+    gender_id, age = age_gender_model.predict(mock_image, bbox_array)
+    assert gender_id in [0, 1], 'Should work with bbox as numpy array'
+    assert 0 <= age <= 120, 'Age should be in valid range'
+
+
+def test_multiple_predictions(age_gender_model, mock_image):
+    bboxes = [
+        [50, 50, 150, 150],
+        [200, 200, 350, 350],
+        [400, 400, 550, 550],
+    ]
+
+    results = []
+    for bbox in bboxes:
+        gender_id, age = age_gender_model.predict(mock_image, bbox)
+        results.append((gender_id, age))
+
+    assert len(results) == 3, 'Should have 3 predictions'
+    for gender_id, age in results:
+        assert gender_id in [0, 1]
+        assert 0 <= age <= 120
+
+
+def test_age_is_positive(age_gender_model, mock_image, mock_bbox):
+    for _ in range(5):
+        gender_id, age = age_gender_model.predict(mock_image, mock_bbox)
+        assert age >= 0, f'Age should be non-negative, got {age}'
+
+
+def test_output_format_for_visualization(age_gender_model, mock_image, mock_bbox):
+    gender_id, age = age_gender_model.predict(mock_image, mock_bbox)
+    gender_str = 'Female' if gender_id == 0 else 'Male'
+    text = f'{gender_str}, {age}y'
+    assert isinstance(text, str), 'Should be able to format as string'
+    assert 'Male' in text or 'Female' in text, 'Text should contain gender'
+    assert 'y' in text, "Text should contain 'y' for years"

tests/test_factory.py

@@ -0,0 +1,274 @@
+import numpy as np
+import pytest
+
+from uniface import (
+    create_detector,
+    create_landmarker,
+    create_recognizer,
+    detect_faces,
+    list_available_detectors,
+)
+from uniface.constants import RetinaFaceWeights, SCRFDWeights
+
+
+# create_detector tests
+def test_create_detector_retinaface():
+    """
+    Test creating a RetinaFace detector using factory function.
+    """
+    detector = create_detector('retinaface')
+    assert detector is not None, 'Failed to create RetinaFace detector'
+
+
+def test_create_detector_scrfd():
+    """
+    Test creating a SCRFD detector using factory function.
+    """
+    detector = create_detector('scrfd')
+    assert detector is not None, 'Failed to create SCRFD detector'
+
+
+def test_create_detector_with_config():
+    """
+    Test creating detector with custom configuration.
+    """
+    detector = create_detector(
+        'retinaface',
+        model_name=RetinaFaceWeights.MNET_V2,
+        conf_thresh=0.8,
+        nms_thresh=0.3,
+    )
+    assert detector is not None, 'Failed to create detector with custom config'
+
+
+def test_create_detector_invalid_method():
+    """
+    Test that invalid detector method raises an error.
+    """
+    with pytest.raises((ValueError, KeyError)):
+        create_detector('invalid_method')
+
+
+def test_create_detector_scrfd_with_model():
+    """
+    Test creating SCRFD detector with specific model.
+    """
+    detector = create_detector('scrfd', model_name=SCRFDWeights.SCRFD_10G_KPS, conf_thresh=0.5)
+    assert detector is not None, 'Failed to create SCRFD with specific model'
+
+
+# create_recognizer tests
+def test_create_recognizer_arcface():
+    """
+    Test creating an ArcFace recognizer using factory function.
+    """
+    recognizer = create_recognizer('arcface')
+    assert recognizer is not None, 'Failed to create ArcFace recognizer'
+
+
+def test_create_recognizer_mobileface():
+    """
+    Test creating a MobileFace recognizer using factory function.
+    """
+    recognizer = create_recognizer('mobileface')
+    assert recognizer is not None, 'Failed to create MobileFace recognizer'
+
+
+def test_create_recognizer_sphereface():
+    """
+    Test creating a SphereFace recognizer using factory function.
+    """
+    recognizer = create_recognizer('sphereface')
+    assert recognizer is not None, 'Failed to create SphereFace recognizer'
+
+
+def test_create_recognizer_invalid_method():
+    """
+    Test that invalid recognizer method raises an error.
+    """
+    with pytest.raises((ValueError, KeyError)):
+        create_recognizer('invalid_method')
+
+
+# create_landmarker tests
+def test_create_landmarker():
+    """
+    Test creating a Landmark106 detector using factory function.
+    """
+    landmarker = create_landmarker('2d106det')
+    assert landmarker is not None, 'Failed to create Landmark106 detector'
+
+
+def test_create_landmarker_default():
+    """
+    Test creating landmarker with default parameters.
+    """
+    landmarker = create_landmarker()
+    assert landmarker is not None, 'Failed to create default landmarker'
+
+
+def test_create_landmarker_invalid_method():
+    """
+    Test that invalid landmarker method raises an error.
+    """
+    with pytest.raises((ValueError, KeyError)):
+        create_landmarker('invalid_method')
+
+
+# detect_faces tests
+def test_detect_faces_retinaface():
+    """
+    Test high-level detect_faces function with RetinaFace.
+    """
+    mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+    faces = detect_faces(mock_image, method='retinaface')
+
+    assert isinstance(faces, list), 'detect_faces should return a list'
+
+
+def test_detect_faces_scrfd():
+    """
+    Test high-level detect_faces function with SCRFD.
+    """
+    mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+    faces = detect_faces(mock_image, method='scrfd')
+
+    assert isinstance(faces, list), 'detect_faces should return a list'
+
+
+def test_detect_faces_with_threshold():
+    """
+    Test detect_faces with custom confidence threshold.
+    """
+    mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+    faces = detect_faces(mock_image, method='retinaface', conf_thresh=0.8)
+
+    assert isinstance(faces, list), 'detect_faces should return a list'
+
+    # All detections should respect threshold
+    for face in faces:
+        assert face['confidence'] >= 0.8, 'All detections should meet confidence threshold'
+
+
+def test_detect_faces_default_method():
+    """
+    Test detect_faces with default method (should use retinaface).
+    """
+    mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+    faces = detect_faces(mock_image)  # No method specified
+
+    assert isinstance(faces, list), 'detect_faces should return a list with default method'
+
+
+def test_detect_faces_empty_image():
+    """
+    Test detect_faces on a blank image.
+    """
+    empty_image = np.zeros((640, 640, 3), dtype=np.uint8)
+    faces = detect_faces(empty_image, method='retinaface')
+
+    assert isinstance(faces, list), 'Should return a list even for empty image'
+    assert len(faces) == 0, 'Should detect no faces in blank image'
+
+
+# list_available_detectors tests
+def test_list_available_detectors():
+    """
+    Test that list_available_detectors returns a dictionary.
+    """
+    detectors = list_available_detectors()
+
+    assert isinstance(detectors, dict), 'Should return a dictionary of detectors'
+    assert len(detectors) > 0, 'Should have at least one detector available'
+
+
+def test_list_available_detectors_contents():
+    """
+    Test that list includes known detectors.
+    """
+    detectors = list_available_detectors()
+
+    # Should include at least these detectors
+    assert 'retinaface' in detectors, "Should include 'retinaface'"
+    assert 'scrfd' in detectors, "Should include 'scrfd'"
+
+
+# Integration tests
+def test_detector_inference_from_factory():
+    """
+    Test that detector created from factory can perform inference.
+    """
+    detector = create_detector('retinaface')
+    mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+
+    faces = detector.detect(mock_image)
+    assert isinstance(faces, list), 'Detector should return list of faces'
+
+
+def test_recognizer_inference_from_factory():
+    """
+    Test that recognizer created from factory can perform inference.
+    """
+    recognizer = create_recognizer('arcface')
+    mock_image = np.random.randint(0, 255, (112, 112, 3), dtype=np.uint8)
+
+    embedding = recognizer.get_embedding(mock_image)
+    assert embedding is not None, 'Recognizer should return embedding'
+    assert embedding.shape[1] == 512, 'Should return 512-dimensional embedding'
+
+
+def test_landmarker_inference_from_factory():
+    """
+    Test that landmarker created from factory can perform inference.
+    """
+    landmarker = create_landmarker('2d106det')
+    mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+    mock_bbox = [100, 100, 300, 300]
+
+    landmarks = landmarker.get_landmarks(mock_image, mock_bbox)
+    assert landmarks is not None, 'Landmarker should return landmarks'
+    assert landmarks.shape == (106, 2), 'Should return 106 landmarks'
+
+
+def test_multiple_detector_creation():
+    """
+    Test that multiple detectors can be created independently.
+    """
+    detector1 = create_detector('retinaface')
+    detector2 = create_detector('scrfd')
+
+    assert detector1 is not None
+    assert detector2 is not None
+    assert detector1 is not detector2, 'Should create separate instances'
+
+
+def test_detector_with_different_configs():
+    """
+    Test creating multiple detectors with different configurations.
+    """
+    detector_high_thresh = create_detector('retinaface', conf_thresh=0.9)
+    detector_low_thresh = create_detector('retinaface', conf_thresh=0.3)
+
+    mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+
+    faces_high = detector_high_thresh.detect(mock_image)
+    faces_low = detector_low_thresh.detect(mock_image)
+
+    # Both should work
+    assert isinstance(faces_high, list)
+    assert isinstance(faces_low, list)
+
+
+def test_factory_returns_correct_types():
+    """
+    Test that factory functions return instances of the correct types.
+    """
+    from uniface import ArcFace, Landmark106, RetinaFace
+
+    detector = create_detector('retinaface')
+    recognizer = create_recognizer('arcface')
+    landmarker = create_landmarker('2d106det')
+
+    assert isinstance(detector, RetinaFace), 'Should return RetinaFace instance'
+    assert isinstance(recognizer, ArcFace), 'Should return ArcFace instance'
+    assert isinstance(landmarker, Landmark106), 'Should return Landmark106 instance'

tests/test_landmark.py

@@ -0,0 +1,107 @@
+import numpy as np
+import pytest
+
+from uniface.landmark import Landmark106
+
+
+@pytest.fixture
+def landmark_model():
+    return Landmark106()
+
+
+@pytest.fixture
+def mock_image():
+    return np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+
+
+@pytest.fixture
+def mock_bbox():
+    return [100, 100, 300, 300]
+
+
+def test_model_initialization(landmark_model):
+    assert landmark_model is not None, 'Landmark106 model initialization failed.'
+
+
+def test_landmark_detection(landmark_model, mock_image, mock_bbox):
+    landmarks = landmark_model.get_landmarks(mock_image, mock_bbox)
+    assert landmarks.shape == (106, 2), f'Expected shape (106, 2), got {landmarks.shape}'
+
+
+def test_landmark_dtype(landmark_model, mock_image, mock_bbox):
+    landmarks = landmark_model.get_landmarks(mock_image, mock_bbox)
+    assert landmarks.dtype == np.float32, f'Expected float32, got {landmarks.dtype}'
+
+
+def test_landmark_coordinates_within_image(landmark_model, mock_image, mock_bbox):
+    landmarks = landmark_model.get_landmarks(mock_image, mock_bbox)
+
+    x_coords = landmarks[:, 0]
+    y_coords = landmarks[:, 1]
+
+    x1, y1, x2, y2 = mock_bbox
+    margin = 50
+
+    x_in_bounds = np.sum((x_coords >= x1 - margin) & (x_coords <= x2 + margin))
+    y_in_bounds = np.sum((y_coords >= y1 - margin) & (y_coords <= y2 + margin))
+
+    assert x_in_bounds >= 95, f'Only {x_in_bounds}/106 x-coordinates within bounds'
+    assert y_in_bounds >= 95, f'Only {y_in_bounds}/106 y-coordinates within bounds'
+
+
+def test_different_bbox_sizes(landmark_model, mock_image):
+    test_bboxes = [
+        [50, 50, 150, 150],
+        [100, 100, 300, 300],
+        [50, 50, 400, 400],
+    ]
+
+    for bbox in test_bboxes:
+        landmarks = landmark_model.get_landmarks(mock_image, bbox)
+        assert landmarks.shape == (106, 2), f'Failed for bbox {bbox}'
+
+
+def test_landmark_array_format(landmark_model, mock_image, mock_bbox):
+    landmarks = landmark_model.get_landmarks(mock_image, mock_bbox)
+    landmarks_int = landmarks.astype(int)
+
+    assert landmarks_int.shape == (106, 2), 'Integer conversion should preserve shape'
+    assert landmarks_int.dtype in [np.int32, np.int64], 'Should convert to integer type'
+
+
+def test_consistency(landmark_model, mock_image, mock_bbox):
+    landmarks1 = landmark_model.get_landmarks(mock_image, mock_bbox)
+    landmarks2 = landmark_model.get_landmarks(mock_image, mock_bbox)
+
+    assert np.allclose(landmarks1, landmarks2), 'Same input should produce same landmarks'
+
+
+def test_different_image_sizes(landmark_model, mock_bbox):
+    test_sizes = [(480, 640, 3), (720, 1280, 3), (1080, 1920, 3)]
+
+    for size in test_sizes:
+        mock_image = np.random.randint(0, 255, size, dtype=np.uint8)
+        landmarks = landmark_model.get_landmarks(mock_image, mock_bbox)
+        assert landmarks.shape == (106, 2), f'Failed for image size {size}'
+
+
+def test_bbox_list_format(landmark_model, mock_image):
+    bbox_list = [100, 100, 300, 300]
+    landmarks = landmark_model.get_landmarks(mock_image, bbox_list)
+    assert landmarks.shape == (106, 2), 'Should work with bbox as list'
+
+
+def test_bbox_array_format(landmark_model, mock_image):
+    bbox_array = np.array([100, 100, 300, 300])
+    landmarks = landmark_model.get_landmarks(mock_image, bbox_array)
+    assert landmarks.shape == (106, 2), 'Should work with bbox as numpy array'
+
+
+def test_landmark_distribution(landmark_model, mock_image, mock_bbox):
+    landmarks = landmark_model.get_landmarks(mock_image, mock_bbox)
+
+    x_variance = np.var(landmarks[:, 0])
+    y_variance = np.var(landmarks[:, 1])
+
+    assert x_variance > 0, 'Landmarks should have variation in x-coordinates'
+    assert y_variance > 0, 'Landmarks should have variation in y-coordinates'

tests/test_parsing.py

@@ -0,0 +1,118 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+import numpy as np
+import pytest
+
+from uniface.constants import ParsingWeights
+from uniface.parsing import BiSeNet, create_face_parser
+
+
+def test_bisenet_initialization():
+    """Test BiSeNet initialization."""
+    parser = BiSeNet()
+    assert parser is not None
+    assert parser.input_size == (512, 512)
+
+
+def test_bisenet_with_different_models():
+    """Test BiSeNet with different model weights."""
+    parser_resnet18 = BiSeNet(model_name=ParsingWeights.RESNET18)
+    parser_resnet34 = BiSeNet(model_name=ParsingWeights.RESNET34)
+
+    assert parser_resnet18 is not None
+    assert parser_resnet34 is not None
+
+
+def test_bisenet_preprocess():
+    """Test preprocessing."""
+    parser = BiSeNet()
+
+    # Create a dummy face image
+    face_image = np.random.randint(0, 255, (256, 256, 3), dtype=np.uint8)
+
+    # Preprocess
+    preprocessed = parser.preprocess(face_image)
+
+    assert preprocessed.shape == (1, 3, 512, 512)
+    assert preprocessed.dtype == np.float32
+
+
+def test_bisenet_postprocess():
+    """Test postprocessing."""
+    parser = BiSeNet()
+
+    # Create dummy model output (batch_size=1, num_classes=19, H=512, W=512)
+    dummy_output = np.random.randn(1, 19, 512, 512).astype(np.float32)
+
+    # Postprocess
+    mask = parser.postprocess(dummy_output, original_size=(256, 256))
+
+    assert mask.shape == (256, 256)
+    assert mask.dtype == np.uint8
+    assert mask.min() >= 0
+    assert mask.max() < 19  # 19 classes (0-18)
+
+
+def test_bisenet_parse():
+    """Test end-to-end parsing."""
+    parser = BiSeNet()
+
+    # Create a dummy face image
+    face_image = np.random.randint(0, 255, (256, 256, 3), dtype=np.uint8)
+
+    # Parse
+    mask = parser.parse(face_image)
+
+    assert mask.shape == (256, 256)
+    assert mask.dtype == np.uint8
+    assert mask.min() >= 0
+    assert mask.max() < 19
+
+
+def test_bisenet_callable():
+    """Test that BiSeNet is callable."""
+    parser = BiSeNet()
+    face_image = np.random.randint(0, 255, (256, 256, 3), dtype=np.uint8)
+
+    # Should work as callable
+    mask = parser(face_image)
+
+    assert mask.shape == (256, 256)
+    assert mask.dtype == np.uint8
+
+
+def test_create_face_parser_with_enum():
+    """Test factory function with enum."""
+    parser = create_face_parser(ParsingWeights.RESNET18)
+    assert parser is not None
+    assert isinstance(parser, BiSeNet)
+
+
+def test_create_face_parser_with_string():
+    """Test factory function with string."""
+    parser = create_face_parser('parsing_resnet18')
+    assert parser is not None
+    assert isinstance(parser, BiSeNet)
+
+
+def test_create_face_parser_invalid_model():
+    """Test factory function with invalid model name."""
+    with pytest.raises(ValueError, match='Unknown face parsing model'):
+        create_face_parser('invalid_model')
+
+
+def test_bisenet_different_input_sizes():
+    """Test parsing with different input image sizes."""
+    parser = BiSeNet()
+
+    # Test with different sizes
+    sizes = [(128, 128), (256, 256), (512, 512), (640, 480)]
+
+    for h, w in sizes:
+        face_image = np.random.randint(0, 255, (h, w, 3), dtype=np.uint8)
+        mask = parser.parse(face_image)
+
+        assert mask.shape == (h, w), f'Failed for size {h}x{w}'
+        assert mask.dtype == np.uint8

tests/test_recognition.py

@@ -0,0 +1,215 @@
+import numpy as np
+import pytest
+
+from uniface.recognition import ArcFace, MobileFace, SphereFace
+
+
+@pytest.fixture
+def arcface_model():
+    """
+    Fixture to initialize the ArcFace model for testing.
+    """
+    return ArcFace()
+
+
+@pytest.fixture
+def mobileface_model():
+    """
+    Fixture to initialize the MobileFace model for testing.
+    """
+    return MobileFace()
+
+
+@pytest.fixture
+def sphereface_model():
+    """
+    Fixture to initialize the SphereFace model for testing.
+    """
+    return SphereFace()
+
+
+@pytest.fixture
+def mock_aligned_face():
+    """
+    Create a mock 112x112 aligned face image.
+    """
+    return np.random.randint(0, 255, (112, 112, 3), dtype=np.uint8)
+
+
+@pytest.fixture
+def mock_landmarks():
+    """
+    Create mock 5-point facial landmarks.
+    """
+    return np.array(
+        [
+            [38.2946, 51.6963],
+            [73.5318, 51.5014],
+            [56.0252, 71.7366],
+            [41.5493, 92.3655],
+            [70.7299, 92.2041],
+        ],
+        dtype=np.float32,
+    )
+
+
+# ArcFace Tests
+def test_arcface_initialization(arcface_model):
+    """
+    Test that the ArcFace model initializes correctly.
+    """
+    assert arcface_model is not None, 'ArcFace model initialization failed.'
+
+
+def test_arcface_embedding_shape(arcface_model, mock_aligned_face):
+    """
+    Test that ArcFace produces embeddings with the correct shape.
+    """
+    embedding = arcface_model.get_embedding(mock_aligned_face)
+
+    # ArcFace typically produces 512-dimensional embeddings
+    assert embedding.shape[1] == 512, f'Expected 512-dim embedding, got {embedding.shape[1]}'
+    assert embedding.shape[0] == 1, 'Embedding should have batch dimension of 1'
+
+
+def test_arcface_normalized_embedding(arcface_model, mock_landmarks):
+    """
+    Test that normalized embeddings have unit length.
+    """
+    # Create a larger mock image for alignment
+    mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+
+    embedding = arcface_model.get_normalized_embedding(mock_image, mock_landmarks)
+
+    # Check that embedding is normalized (L2 norm ≈ 1.0)
+    norm = np.linalg.norm(embedding)
+    assert np.isclose(norm, 1.0, atol=1e-5), f'Normalized embedding should have norm 1.0, got {norm}'
+
+
+def test_arcface_embedding_dtype(arcface_model, mock_aligned_face):
+    """
+    Test that embeddings have the correct data type.
+    """
+    embedding = arcface_model.get_embedding(mock_aligned_face)
+    assert embedding.dtype == np.float32, f'Expected float32, got {embedding.dtype}'
+
+
+def test_arcface_consistency(arcface_model, mock_aligned_face):
+    """
+    Test that the same input produces the same embedding.
+    """
+    embedding1 = arcface_model.get_embedding(mock_aligned_face)
+    embedding2 = arcface_model.get_embedding(mock_aligned_face)
+
+    assert np.allclose(embedding1, embedding2), 'Same input should produce same embedding'
+
+
+# MobileFace Tests
+def test_mobileface_initialization(mobileface_model):
+    """
+    Test that the MobileFace model initializes correctly.
+    """
+    assert mobileface_model is not None, 'MobileFace model initialization failed.'
+
+
+def test_mobileface_embedding_shape(mobileface_model, mock_aligned_face):
+    """
+    Test that MobileFace produces embeddings with the correct shape.
+    """
+    embedding = mobileface_model.get_embedding(mock_aligned_face)
+
+    # MobileFace typically produces 512-dimensional embeddings
+    assert embedding.shape[1] == 512, f'Expected 512-dim embedding, got {embedding.shape[1]}'
+    assert embedding.shape[0] == 1, 'Embedding should have batch dimension of 1'
+
+
+def test_mobileface_normalized_embedding(mobileface_model, mock_landmarks):
+    """
+    Test that MobileFace normalized embeddings have unit length.
+    """
+    mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+
+    embedding = mobileface_model.get_normalized_embedding(mock_image, mock_landmarks)
+
+    norm = np.linalg.norm(embedding)
+    assert np.isclose(norm, 1.0, atol=1e-5), f'Normalized embedding should have norm 1.0, got {norm}'
+
+
+# SphereFace Tests
+def test_sphereface_initialization(sphereface_model):
+    """
+    Test that the SphereFace model initializes correctly.
+    """
+    assert sphereface_model is not None, 'SphereFace model initialization failed.'
+
+
+def test_sphereface_embedding_shape(sphereface_model, mock_aligned_face):
+    """
+    Test that SphereFace produces embeddings with the correct shape.
+    """
+    embedding = sphereface_model.get_embedding(mock_aligned_face)
+
+    # SphereFace typically produces 512-dimensional embeddings
+    assert embedding.shape[1] == 512, f'Expected 512-dim embedding, got {embedding.shape[1]}'
+    assert embedding.shape[0] == 1, 'Embedding should have batch dimension of 1'
+
+
+def test_sphereface_normalized_embedding(sphereface_model, mock_landmarks):
+    """
+    Test that SphereFace normalized embeddings have unit length.
+    """
+    mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+
+    embedding = sphereface_model.get_normalized_embedding(mock_image, mock_landmarks)
+
+    norm = np.linalg.norm(embedding)
+    assert np.isclose(norm, 1.0, atol=1e-5), f'Normalized embedding should have norm 1.0, got {norm}'
+
+
+# Cross-model comparison tests
+def test_different_models_different_embeddings(arcface_model, mobileface_model, mock_aligned_face):
+    """
+    Test that different models produce different embeddings for the same input.
+    """
+    arcface_emb = arcface_model.get_embedding(mock_aligned_face)
+    mobileface_emb = mobileface_model.get_embedding(mock_aligned_face)
+
+    # Embeddings should be different (with high probability for random input)
+    # We check that they're not identical
+    assert not np.allclose(arcface_emb, mobileface_emb), 'Different models should produce different embeddings'
+
+
+def test_embedding_similarity_computation(arcface_model, mock_aligned_face):
+    """
+    Test computing similarity between embeddings.
+    """
+    # Get two embeddings
+    emb1 = arcface_model.get_embedding(mock_aligned_face)
+
+    # Create a slightly different image
+    mock_aligned_face2 = mock_aligned_face.copy()
+    mock_aligned_face2[:10, :10] = 0  # Modify a small region
+    emb2 = arcface_model.get_embedding(mock_aligned_face2)
+
+    # Compute cosine similarity
+    from uniface import compute_similarity
+
+    similarity = compute_similarity(emb1, emb2)
+
+    # Similarity should be between -1 and 1
+    assert -1.0 <= similarity <= 1.0, f'Similarity should be in [-1, 1], got {similarity}'
+
+
+def test_same_face_high_similarity(arcface_model, mock_aligned_face):
+    """
+    Test that the same face produces high similarity.
+    """
+    emb1 = arcface_model.get_embedding(mock_aligned_face)
+    emb2 = arcface_model.get_embedding(mock_aligned_face)
+
+    from uniface import compute_similarity
+
+    similarity = compute_similarity(emb1, emb2)
+
+    # Same image should have similarity close to 1.0
+    assert similarity > 0.99, f'Same face should have similarity > 0.99, got {similarity}'

tests/test_retinaface.py

@@ -1,15 +1,14 @@
-import pytest
 import numpy as np
-from uniface import RetinaFace
+import pytest
+
+from uniface.constants import RetinaFaceWeights
+from uniface.detection import RetinaFace
 
 
 @pytest.fixture
 def retinaface_model():
-    """
-    Fixture to initialize the RetinaFace model for testing.
-    """
     return RetinaFace(
-        model="retinaface_mnet_v2",
+        model_name=RetinaFaceWeights.MNET_V2,
         conf_thresh=0.5,
         pre_nms_topk=5000,
         nms_thresh=0.4,
@@ -18,61 +17,39 @@ def retinaface_model():
 
 
 def test_model_initialization(retinaface_model):
-    """
-    Test that the RetinaFace model initializes correctly.
-    """
-    assert retinaface_model is not None, "Model initialization failed."
+    assert retinaface_model is not None, 'Model initialization failed.'
 
 
 def test_inference_on_640x640_image(retinaface_model):
-    """
-    Test inference on a 640x640 BGR image.
-    """
-    # Generate a mock 640x640 BGR image
     mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+    faces = retinaface_model.detect(mock_image)
 
-    # Run inference
-    detections, landmarks = retinaface_model.detect(mock_image)
+    assert isinstance(faces, list), 'Detections should be a list.'
 
-    # Check output types
-    assert isinstance(detections, np.ndarray), "Detections should be a numpy array."
-    assert isinstance(landmarks, np.ndarray), "Landmarks should be a numpy array."
+    for face in faces:
+        assert isinstance(face, dict), 'Each detection should be a dictionary.'
+        assert 'bbox' in face, "Each detection should have a 'bbox' key."
+        assert 'confidence' in face, "Each detection should have a 'confidence' key."
+        assert 'landmarks' in face, "Each detection should have a 'landmarks' key."
 
-    # Check that detections have the expected shape
-    if detections.size > 0:  # If faces are detected
-        assert detections.shape[1] == 5, "Each detection should have 5 values (x1, y1, x2, y2, score)."
+        bbox = face['bbox']
+        assert len(bbox) == 4, 'BBox should have 4 values (x1, y1, x2, y2).'
 
-    # Check landmarks shape
-    if landmarks.size > 0:
-        assert landmarks.shape[1:] == (5, 2), "Landmarks should have shape (N, 5, 2)."
+        landmarks = face['landmarks']
+        assert len(landmarks) == 5, 'Should have 5 landmark points.'
+        assert all(len(pt) == 2 for pt in landmarks), 'Each landmark should be (x, y).'
 
 
 def test_confidence_threshold(retinaface_model):
-    """
-    Test that detections respect the confidence threshold.
-    """
-    # Generate a mock 640x640 BGR image
     mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+    faces = retinaface_model.detect(mock_image)
 
-    # Run inference
-    detections, _ = retinaface_model.detect(mock_image)
-
-    # Ensure all detections have confidence scores above the threshold
-    if detections.size > 0:  # If faces are detected
-        confidence_scores = detections[:, 4]
-        assert (confidence_scores >= 0.5).all(), "Some detections have confidence below the threshold."
+    for face in faces:
+        confidence = face['confidence']
+        assert confidence >= 0.5, f'Detection has confidence {confidence} below threshold 0.5'
 
 
 def test_no_faces_detected(retinaface_model):
-    """
-    Test inference on an image without detectable faces.
-    """
-    # Generate an empty (black) 640x640 image
     empty_image = np.zeros((640, 640, 3), dtype=np.uint8)
-
-    # Run inference
-    detections, landmarks = retinaface_model.detect(empty_image)
-
-    # Ensure no detections or landmarks are found
-    assert detections.size == 0, "Detections should be empty for a blank image."
-    assert landmarks.size == 0, "Landmarks should be empty for a blank image."
+    faces = retinaface_model.detect(empty_image)
+    assert len(faces) == 0, 'Should detect no faces in a blank image.'

tests/test_scrfd.py

@@ -0,0 +1,71 @@
+import numpy as np
+import pytest
+
+from uniface.constants import SCRFDWeights
+from uniface.detection import SCRFD
+
+
+@pytest.fixture
+def scrfd_model():
+    return SCRFD(
+        model_name=SCRFDWeights.SCRFD_500M_KPS,
+        conf_thresh=0.5,
+        nms_thresh=0.4,
+    )
+
+
+def test_model_initialization(scrfd_model):
+    assert scrfd_model is not None, 'Model initialization failed.'
+
+
+def test_inference_on_640x640_image(scrfd_model):
+    mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+    faces = scrfd_model.detect(mock_image)
+
+    assert isinstance(faces, list), 'Detections should be a list.'
+
+    for face in faces:
+        assert isinstance(face, dict), 'Each detection should be a dictionary.'
+        assert 'bbox' in face, "Each detection should have a 'bbox' key."
+        assert 'confidence' in face, "Each detection should have a 'confidence' key."
+        assert 'landmarks' in face, "Each detection should have a 'landmarks' key."
+
+        bbox = face['bbox']
+        assert len(bbox) == 4, 'BBox should have 4 values (x1, y1, x2, y2).'
+
+        landmarks = face['landmarks']
+        assert len(landmarks) == 5, 'Should have 5 landmark points.'
+        assert all(len(pt) == 2 for pt in landmarks), 'Each landmark should be (x, y).'
+
+
+def test_confidence_threshold(scrfd_model):
+    mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+    faces = scrfd_model.detect(mock_image)
+
+    for face in faces:
+        confidence = face['confidence']
+        assert confidence >= 0.5, f'Detection has confidence {confidence} below threshold 0.5'
+
+
+def test_no_faces_detected(scrfd_model):
+    empty_image = np.zeros((640, 640, 3), dtype=np.uint8)
+    faces = scrfd_model.detect(empty_image)
+    assert len(faces) == 0, 'Should detect no faces in a blank image.'
+
+
+def test_different_input_sizes(scrfd_model):
+    test_sizes = [(480, 640, 3), (720, 1280, 3), (1080, 1920, 3)]
+
+    for size in test_sizes:
+        mock_image = np.random.randint(0, 255, size, dtype=np.uint8)
+        faces = scrfd_model.detect(mock_image)
+        assert isinstance(faces, list), f'Should return list for size {size}'
+
+
+def test_scrfd_10g_model():
+    model = SCRFD(model_name=SCRFDWeights.SCRFD_10G_KPS, conf_thresh=0.5)
+    assert model is not None, 'SCRFD 10G model initialization failed.'
+
+    mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+    faces = model.detect(mock_image)
+    assert isinstance(faces, list), 'SCRFD 10G should return list of detections.'

tests/test_utils.py

@@ -0,0 +1,262 @@
+import numpy as np
+import pytest
+
+from uniface import compute_similarity, face_alignment
+
+
+@pytest.fixture
+def mock_image():
+    """
+    Create a mock 640x640 BGR image.
+    """
+    return np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+
+
+@pytest.fixture
+def mock_landmarks():
+    """
+    Create mock 5-point facial landmarks.
+    Standard positions for a face roughly centered at (112/2, 112/2).
+    """
+    return np.array(
+        [
+            [38.2946, 51.6963],  # Left eye
+            [73.5318, 51.5014],  # Right eye
+            [56.0252, 71.7366],  # Nose
+            [41.5493, 92.3655],  # Left mouth corner
+            [70.7299, 92.2041],  # Right mouth corner
+        ],
+        dtype=np.float32,
+    )
+
+
+# compute_similarity tests
+def test_compute_similarity_same_embedding():
+    """
+    Test that similarity of an embedding with itself is 1.0.
+    """
+    embedding = np.random.randn(1, 512).astype(np.float32)
+    embedding = embedding / np.linalg.norm(embedding)  # Normalize
+
+    similarity = compute_similarity(embedding, embedding)
+    assert np.isclose(similarity, 1.0, atol=1e-5), f'Self-similarity should be 1.0, got {similarity}'
+
+
+def test_compute_similarity_range():
+    """
+    Test that similarity is always in the range [-1, 1].
+    """
+    # Test with multiple random embeddings
+    for _ in range(10):
+        emb1 = np.random.randn(1, 512).astype(np.float32)
+        emb2 = np.random.randn(1, 512).astype(np.float32)
+
+        # Normalize
+        emb1 = emb1 / np.linalg.norm(emb1)
+        emb2 = emb2 / np.linalg.norm(emb2)
+
+        similarity = compute_similarity(emb1, emb2)
+        assert -1.0 <= similarity <= 1.0, f'Similarity should be in [-1, 1], got {similarity}'
+
+
+def test_compute_similarity_orthogonal():
+    """
+    Test that orthogonal embeddings have similarity close to 0.
+    """
+    # Create orthogonal embeddings
+    emb1 = np.zeros((1, 512), dtype=np.float32)
+    emb1[0, 0] = 1.0  # [1, 0, 0, ..., 0]
+
+    emb2 = np.zeros((1, 512), dtype=np.float32)
+    emb2[0, 1] = 1.0  # [0, 1, 0, ..., 0]
+
+    similarity = compute_similarity(emb1, emb2)
+    assert np.isclose(similarity, 0.0, atol=1e-5), f'Orthogonal embeddings should have similarity 0.0, got {similarity}'
+
+
+def test_compute_similarity_opposite():
+    """
+    Test that opposite embeddings have similarity close to -1.
+    """
+    emb1 = np.ones((1, 512), dtype=np.float32)
+    emb1 = emb1 / np.linalg.norm(emb1)
+
+    emb2 = -emb1  # Opposite direction
+
+    similarity = compute_similarity(emb1, emb2)
+    assert np.isclose(similarity, -1.0, atol=1e-5), f'Opposite embeddings should have similarity -1.0, got {similarity}'
+
+
+def test_compute_similarity_symmetry():
+    """
+    Test that similarity(A, B) == similarity(B, A).
+    """
+    emb1 = np.random.randn(1, 512).astype(np.float32)
+    emb2 = np.random.randn(1, 512).astype(np.float32)
+
+    # Normalize
+    emb1 = emb1 / np.linalg.norm(emb1)
+    emb2 = emb2 / np.linalg.norm(emb2)
+
+    sim_12 = compute_similarity(emb1, emb2)
+    sim_21 = compute_similarity(emb2, emb1)
+
+    assert np.isclose(sim_12, sim_21), 'Similarity should be symmetric'
+
+
+def test_compute_similarity_dtype():
+    """
+    Test that compute_similarity returns a float.
+    """
+    emb1 = np.random.randn(1, 512).astype(np.float32)
+    emb2 = np.random.randn(1, 512).astype(np.float32)
+
+    # Normalize
+    emb1 = emb1 / np.linalg.norm(emb1)
+    emb2 = emb2 / np.linalg.norm(emb2)
+
+    similarity = compute_similarity(emb1, emb2)
+    assert isinstance(similarity, (float, np.floating)), f'Similarity should be float, got {type(similarity)}'
+
+
+# face_alignment tests
+def test_face_alignment_output_shape(mock_image, mock_landmarks):
+    """
+    Test that face_alignment produces output with the correct shape.
+    """
+    aligned, _ = face_alignment(mock_image, mock_landmarks, image_size=(112, 112))
+
+    assert aligned.shape == (112, 112, 3), f'Expected shape (112, 112, 3), got {aligned.shape}'
+
+
+def test_face_alignment_dtype(mock_image, mock_landmarks):
+    """
+    Test that aligned face has the correct data type.
+    """
+    aligned, _ = face_alignment(mock_image, mock_landmarks, image_size=(112, 112))
+
+    assert aligned.dtype == np.uint8, f'Expected uint8, got {aligned.dtype}'
+
+
+def test_face_alignment_different_sizes(mock_image, mock_landmarks):
+    """
+    Test face alignment with different output sizes.
+    """
+    # Only test sizes that are multiples of 112 or 128 as required by the function
+    test_sizes = [(112, 112), (128, 128), (224, 224)]
+
+    for size in test_sizes:
+        aligned, _ = face_alignment(mock_image, mock_landmarks, image_size=size)
+        assert aligned.shape == (*size, 3), f'Failed for size {size}'
+
+
+def test_face_alignment_consistency(mock_image, mock_landmarks):
+    """
+    Test that the same input produces the same aligned face.
+    """
+    aligned1, _ = face_alignment(mock_image, mock_landmarks, image_size=(112, 112))
+    aligned2, _ = face_alignment(mock_image, mock_landmarks, image_size=(112, 112))
+
+    assert np.allclose(aligned1, aligned2), 'Same input should produce same aligned face'
+
+
+def test_face_alignment_landmarks_as_list(mock_image):
+    """
+    Test that landmarks can be passed as a list of lists (converted to array).
+    """
+    landmarks_list = [
+        [38.2946, 51.6963],
+        [73.5318, 51.5014],
+        [56.0252, 71.7366],
+        [41.5493, 92.3655],
+        [70.7299, 92.2041],
+    ]
+
+    # Convert list to numpy array before passing to face_alignment
+    landmarks_array = np.array(landmarks_list, dtype=np.float32)
+    aligned, _ = face_alignment(mock_image, landmarks_array, image_size=(112, 112))
+    assert aligned.shape == (112, 112, 3), 'Should work with landmarks as array'
+
+
+def test_face_alignment_value_range(mock_image, mock_landmarks):
+    """
+    Test that aligned face pixel values are in valid range [0, 255].
+    """
+    aligned, _ = face_alignment(mock_image, mock_landmarks, image_size=(112, 112))
+
+    assert np.all(aligned >= 0), 'Pixel values should be >= 0'
+    assert np.all(aligned <= 255), 'Pixel values should be <= 255'
+
+
+def test_face_alignment_not_all_zeros(mock_image, mock_landmarks):
+    """
+    Test that aligned face is not all zeros (actual transformation occurred).
+    """
+    aligned, _ = face_alignment(mock_image, mock_landmarks, image_size=(112, 112))
+
+    # At least some pixels should be non-zero
+    assert np.any(aligned > 0), 'Aligned face should have some non-zero pixels'
+
+
+def test_face_alignment_from_different_positions(mock_image):
+    """
+    Test alignment with landmarks at different positions in the image.
+    """
+    # Landmarks at different positions
+    positions = [
+        np.array(
+            [[100, 100], [150, 100], [125, 130], [110, 150], [140, 150]],
+            dtype=np.float32,
+        ),
+        np.array(
+            [[300, 200], [350, 200], [325, 230], [310, 250], [340, 250]],
+            dtype=np.float32,
+        ),
+        np.array(
+            [[500, 400], [550, 400], [525, 430], [510, 450], [540, 450]],
+            dtype=np.float32,
+        ),
+    ]
+
+    for landmarks in positions:
+        aligned, _ = face_alignment(mock_image, landmarks, image_size=(112, 112))
+        assert aligned.shape == (112, 112, 3), f'Failed for landmarks at {landmarks[0]}'
+
+
+def test_face_alignment_landmark_count(mock_image):
+    """
+    Test that face_alignment works specifically with 5-point landmarks.
+    """
+    # Standard 5-point landmarks
+    landmarks_5pt = np.array(
+        [
+            [38.2946, 51.6963],
+            [73.5318, 51.5014],
+            [56.0252, 71.7366],
+            [41.5493, 92.3655],
+            [70.7299, 92.2041],
+        ],
+        dtype=np.float32,
+    )
+
+    aligned, _ = face_alignment(mock_image, landmarks_5pt, image_size=(112, 112))
+    assert aligned.shape == (112, 112, 3), 'Should work with 5-point landmarks'
+
+
+def test_compute_similarity_with_recognition_embeddings():
+    """
+    Test compute_similarity with realistic embedding dimensions.
+    """
+    # Simulate ArcFace/MobileFace/SphereFace embeddings (512-dim)
+    emb1 = np.random.randn(1, 512).astype(np.float32)
+    emb2 = np.random.randn(1, 512).astype(np.float32)
+
+    # Normalize (as done in get_normalized_embedding)
+    emb1 = emb1 / np.linalg.norm(emb1)
+    emb2 = emb2 / np.linalg.norm(emb2)
+
+    similarity = compute_similarity(emb1, emb2)
+
+    # Should be a valid similarity score
+    assert -1.0 <= similarity <= 1.0
+    assert isinstance(similarity, (float, np.floating))

uniface/__init__.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Yakhyokhuja Valikhujaev
+# Copyright 2025 Yakhyokhuja Valikhujaev
 #
 # Licensed under the MIT License.
 # You may obtain a copy of the License at
@@ -11,18 +11,75 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+__license__ = 'MIT'
+__author__ = 'Yakhyokhuja Valikhujaev'
+__version__ = '1.5.0'
 
-from uniface.retinaface import RetinaFace
-from uniface.log import Logger
+
+from uniface.face_utils import compute_similarity, face_alignment
+from uniface.log import Logger, enable_logging
 from uniface.model_store import verify_model_weights
-from uniface.version import __version__, __author__
-from uniface.visualization import draw_detections
+from uniface.visualization import draw_detections, vis_parsing_maps
+
+from .analyzer import FaceAnalyzer
+from .attribute import AgeGender
+from .face import Face
+
+try:
+    from .attribute import Emotion
+except ImportError:
+    Emotion = None  # PyTorch not installed
+from .detection import (
+    SCRFD,
+    RetinaFace,
+    YOLOv5Face,
+    create_detector,
+    detect_faces,
+    list_available_detectors,
+)
+from .gaze import MobileGaze, create_gaze_estimator
+from .landmark import Landmark106, create_landmarker
+from .parsing import BiSeNet, create_face_parser
+from .recognition import ArcFace, MobileFace, SphereFace, create_recognizer
 
 __all__ = [
-    "__version__",
-    "__author__"
-    "RetinaFace",
-    "Logger",
-    "verify_model_weights",
-    "draw_detections"
+    '__author__',
+    '__license__',
+    '__version__',
+    # Core classes
+    'Face',
+    'FaceAnalyzer',
+    # Factory functions
+    'create_detector',
+    'create_face_parser',
+    'create_gaze_estimator',
+    'create_landmarker',
+    'create_recognizer',
+    'detect_faces',
+    'list_available_detectors',
+    # Detection models
+    'RetinaFace',
+    'SCRFD',
+    'YOLOv5Face',
+    # Recognition models
+    'ArcFace',
+    'MobileFace',
+    'SphereFace',
+    # Landmark models
+    'Landmark106',
+    # Gaze models
+    'MobileGaze',
+    # Parsing models
+    'BiSeNet',
+    # Attribute models
+    'AgeGender',
+    'Emotion',
+    # Utilities
+    'compute_similarity',
+    'draw_detections',
+    'vis_parsing_maps',
+    'face_alignment',
+    'verify_model_weights',
+    'Logger',
+    'enable_logging',
 ]

uniface/analyzer.py

@@ -0,0 +1,83 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from typing import List, Optional
+
+import numpy as np
+
+from uniface.attribute.age_gender import AgeGender
+from uniface.detection.base import BaseDetector
+from uniface.face import Face
+from uniface.log import Logger
+from uniface.recognition.base import BaseRecognizer
+
+__all__ = ['FaceAnalyzer']
+
+
+class FaceAnalyzer:
+    """Unified face analyzer combining detection, recognition, and attributes."""
+
+    def __init__(
+        self,
+        detector: BaseDetector,
+        recognizer: Optional[BaseRecognizer] = None,
+        age_gender: Optional[AgeGender] = None,
+    ) -> None:
+        self.detector = detector
+        self.recognizer = recognizer
+        self.age_gender = age_gender
+
+        Logger.info(f'Initialized FaceAnalyzer with detector={detector.__class__.__name__}')
+        if recognizer:
+            Logger.info(f'  - Recognition enabled: {recognizer.__class__.__name__}')
+        if age_gender:
+            Logger.info(f'  - Age/Gender enabled: {age_gender.__class__.__name__}')
+
+    def analyze(self, image: np.ndarray) -> List[Face]:
+        """Analyze faces in an image."""
+        detections = self.detector.detect(image)
+        Logger.debug(f'Detected {len(detections)} face(s)')
+
+        faces = []
+        for idx, detection in enumerate(detections):
+            bbox = detection['bbox']
+            confidence = detection['confidence']
+            landmarks = detection['landmarks']
+
+            embedding = None
+            if self.recognizer is not None:
+                try:
+                    embedding = self.recognizer.get_normalized_embedding(image, landmarks)
+                    Logger.debug(f'  Face {idx + 1}: Extracted embedding with shape {embedding.shape}')
+                except Exception as e:
+                    Logger.warning(f'  Face {idx + 1}: Failed to extract embedding: {e}')
+
+            age, gender = None, None
+            if self.age_gender is not None:
+                try:
+                    gender, age = self.age_gender.predict(image, bbox)
+                    Logger.debug(f'  Face {idx + 1}: Age={age}, Gender={gender}')
+                except Exception as e:
+                    Logger.warning(f'  Face {idx + 1}: Failed to predict age/gender: {e}')
+
+            face = Face(
+                bbox=bbox,
+                confidence=confidence,
+                landmarks=landmarks,
+                embedding=embedding,
+                age=age,
+                gender=gender,
+            )
+            faces.append(face)
+
+        Logger.info(f'Analysis complete: {len(faces)} face(s) processed')
+        return faces
+
+    def __repr__(self) -> str:
+        parts = [f'FaceAnalyzer(detector={self.detector.__class__.__name__}']
+        if self.recognizer:
+            parts.append(f'recognizer={self.recognizer.__class__.__name__}')
+        if self.age_gender:
+            parts.append(f'age_gender={self.age_gender.__class__.__name__}')
+        return ', '.join(parts) + ')'

uniface/attribute/__init__.py

@@ -0,0 +1,99 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from typing import Any, Dict, List, Union
+
+import numpy as np
+
+from uniface.attribute.age_gender import AgeGender
+from uniface.attribute.base import Attribute
+from uniface.constants import AgeGenderWeights, DDAMFNWeights
+
+# Emotion requires PyTorch - make it optional
+try:
+    from uniface.attribute.emotion import Emotion
+
+    _EMOTION_AVAILABLE = True
+except ImportError:
+    Emotion = None
+    _EMOTION_AVAILABLE = False
+
+# Public API for the attribute module
+__all__ = ['AgeGender', 'Emotion', 'create_attribute_predictor', 'predict_attributes']
+
+# A mapping from model enums to their corresponding attribute classes
+_ATTRIBUTE_MODELS = {
+    **{model: AgeGender for model in AgeGenderWeights},
+}
+
+# Add Emotion models only if PyTorch is available
+if _EMOTION_AVAILABLE:
+    _ATTRIBUTE_MODELS.update({model: Emotion for model in DDAMFNWeights})
+
+
+def create_attribute_predictor(model_name: Union[AgeGenderWeights, DDAMFNWeights], **kwargs: Any) -> Attribute:
+    """
+    Factory function to create an attribute predictor instance.
+
+    This high-level API simplifies the creation of attribute models by
+    dynamically selecting the correct class based on the provided model enum.
+
+    Args:
+        model_name: The enum corresponding to the desired attribute model
+                    (e.g., AgeGenderWeights.DEFAULT or DDAMFNWeights.AFFECNET7).
+        **kwargs: Additional keyword arguments to pass to the model's constructor.
+
+    Returns:
+        An initialized instance of an Attribute predictor class (e.g., AgeGender).
+
+    Raises:
+        ValueError: If the provided model_name is not a supported enum.
+    """
+    model_class = _ATTRIBUTE_MODELS.get(model_name)
+
+    if model_class is None:
+        raise ValueError(
+            f'Unsupported attribute model: {model_name}. Please choose from AgeGenderWeights or DDAMFNWeights.'
+        )
+
+    # Pass model_name to the constructor, as some classes might need it
+    return model_class(model_name=model_name, **kwargs)
+
+
+def predict_attributes(
+    image: np.ndarray, detections: List[Dict[str, np.ndarray]], predictor: Attribute
+) -> List[Dict[str, Any]]:
+    """
+    High-level API to predict attributes for multiple detected faces.
+
+    This function iterates through a list of face detections, runs the
+    specified attribute predictor on each one, and appends the results back
+    into the detection dictionary.
+
+    Args:
+        image (np.ndarray): The full input image in BGR format.
+        detections (List[Dict]): A list of detection results, where each dict
+                                 must contain a 'bbox' and optionally 'landmark'.
+        predictor (Attribute): An initialized attribute predictor instance,
+                               created by `create_attribute_predictor`.
+
+    Returns:
+        The list of detections, where each dictionary is updated with a new
+        'attributes' key containing the prediction result.
+    """
+    for face in detections:
+        # Initialize attributes dict if it doesn't exist
+        if 'attributes' not in face:
+            face['attributes'] = {}
+
+        if isinstance(predictor, AgeGender):
+            gender_id, age = predictor(image, face['bbox'])
+            face['attributes']['gender_id'] = gender_id
+            face['attributes']['age'] = age
+        elif isinstance(predictor, Emotion):
+            emotion, confidence = predictor(image, face['landmark'])
+            face['attributes']['emotion'] = emotion
+            face['attributes']['confidence'] = confidence
+
+    return detections

uniface/attribute/age_gender.py

@@ -0,0 +1,211 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from typing import List, Optional, Tuple, Union
+
+import cv2
+import numpy as np
+
+from uniface.attribute.base import Attribute
+from uniface.constants import AgeGenderWeights
+from uniface.face_utils import bbox_center_alignment
+from uniface.log import Logger
+from uniface.model_store import verify_model_weights
+from uniface.onnx_utils import create_onnx_session
+
+__all__ = ['AgeGender']
+
+
+class AgeGender(Attribute):
+    """
+    Age and gender prediction model using ONNX Runtime.
+
+    This class inherits from the base `Attribute` class and implements the
+    functionality for predicting age (in years) and gender ID (0 for Female,
+    1 for Male) from a face image. It requires a bounding box to locate the face.
+
+    Args:
+        model_name (AgeGenderWeights): The enum specifying the model weights to load.
+            Defaults to `AgeGenderWeights.DEFAULT`.
+        input_size (Optional[Tuple[int, int]]): Input size (height, width).
+            If None, automatically detected from model metadata. Defaults to None.
+    """
+
+    def __init__(
+        self,
+        model_name: AgeGenderWeights = AgeGenderWeights.DEFAULT,
+        input_size: Optional[Tuple[int, int]] = None,
+    ) -> None:
+        """
+        Initializes the AgeGender prediction model.
+
+        Args:
+            model_name (AgeGenderWeights): The enum specifying the model weights to load.
+            input_size (Optional[Tuple[int, int]]): Input size (height, width).
+                If None, automatically detected from model metadata. Defaults to None.
+        """
+        Logger.info(f'Initializing AgeGender with model={model_name.name}')
+        self.model_path = verify_model_weights(model_name)
+        self._user_input_size = input_size  # Store user preference
+        self._initialize_model()
+
+    def _initialize_model(self) -> None:
+        """
+        Initializes the ONNX model and creates an inference session.
+        """
+        try:
+            self.session = create_onnx_session(self.model_path)
+            # Get model input details from the loaded model
+            input_meta = self.session.get_inputs()[0]
+            self.input_name = input_meta.name
+
+            # Use user-provided size if given, otherwise auto-detect from model
+            model_input_size = tuple(input_meta.shape[2:4])  # (height, width)
+            if self._user_input_size is not None:
+                self.input_size = self._user_input_size
+                if self._user_input_size != model_input_size:
+                    Logger.warning(
+                        f'Using custom input_size {self.input_size}, '
+                        f'but model expects {model_input_size}. This may affect accuracy.'
+                    )
+            else:
+                self.input_size = model_input_size
+
+            self.output_names = [output.name for output in self.session.get_outputs()]
+            Logger.info(f'Successfully initialized AgeGender model with input size {self.input_size}')
+        except Exception as e:
+            Logger.error(
+                f"Failed to load AgeGender model from '{self.model_path}'",
+                exc_info=True,
+            )
+            raise RuntimeError(f'Failed to initialize AgeGender model: {e}') from e
+
+    def preprocess(self, image: np.ndarray, bbox: Union[List, np.ndarray]) -> np.ndarray:
+        """
+        Aligns the face based on the bounding box and preprocesses it for inference.
+
+        Args:
+            image (np.ndarray): The full input image in BGR format.
+            bbox (Union[List, np.ndarray]): The face bounding box coordinates [x1, y1, x2, y2].
+
+        Returns:
+            np.ndarray: The preprocessed image blob ready for inference.
+        """
+        bbox = np.asarray(bbox)
+
+        width, height = bbox[2] - bbox[0], bbox[3] - bbox[1]
+        center = ((bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2)
+        scale = self.input_size[1] / (max(width, height) * 1.5)
+
+        # **Rotation parameter restored here**
+        rotation = 0.0
+        aligned_face, _ = bbox_center_alignment(image, center, self.input_size[1], scale, rotation)
+
+        blob = cv2.dnn.blobFromImage(
+            aligned_face,
+            scalefactor=1.0,
+            size=self.input_size[::-1],
+            mean=(0.0, 0.0, 0.0),
+            swapRB=True,
+        )
+        return blob
+
+    def postprocess(self, prediction: np.ndarray) -> Tuple[int, int]:
+        """
+        Processes the raw model output to extract gender and age.
+
+        Args:
+            prediction (np.ndarray): The raw output from the model inference.
+
+        Returns:
+            Tuple[int, int]: A tuple containing the predicted gender ID (0 for Female, 1 for Male)
+                             and age (in years).
+        """
+        # First two values are gender logits
+        gender_id = int(np.argmax(prediction[:2]))
+        # Third value is normalized age, scaled by 100
+        age = int(np.round(prediction[2] * 100))
+        return gender_id, age
+
+    def predict(self, image: np.ndarray, bbox: Union[List, np.ndarray]) -> Tuple[int, int]:
+        """
+        Predicts age and gender for a single face specified by a bounding box.
+
+        Args:
+            image (np.ndarray): The full input image in BGR format.
+            bbox (Union[List, np.ndarray]): The face bounding box coordinates [x1, y1, x2, y2].
+
+        Returns:
+            Tuple[int, int]: A tuple containing the predicted gender ID (0 for Female, 1 for Male) and age.
+        """
+        face_blob = self.preprocess(image, bbox)
+        prediction = self.session.run(self.output_names, {self.input_name: face_blob})[0][0]
+        gender_id, age = self.postprocess(prediction)
+        return gender_id, age
+
+
+# TODO: below is only for testing, remove it later
+if __name__ == '__main__':
+    # To run this script, you need to have uniface.detection installed
+    # or available in your path.
+    from uniface.constants import RetinaFaceWeights
+    from uniface.detection import create_detector
+
+    print('Initializing models for live inference...')
+    # 1. Initialize the face detector
+    # Using a smaller model for faster real-time performance
+    detector = create_detector(model_name=RetinaFaceWeights.MNET_V2)
+
+    # 2. Initialize the attribute predictor
+    age_gender_predictor = AgeGender()
+
+    # 3. Start webcam capture
+    cap = cv2.VideoCapture(0)
+    if not cap.isOpened():
+        print('Error: Could not open webcam.')
+        exit()
+
+    print("Starting webcam feed. Press 'q' to quit.")
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            print('Error: Failed to capture frame.')
+            break
+
+        # Detect faces in the current frame
+        detections = detector.detect(frame)
+
+        # For each detected face, predict age and gender
+        for detection in detections:
+            box = detection['bbox']
+            x1, y1, x2, y2 = map(int, box)
+
+            # Predict attributes
+            gender_id, age = age_gender_predictor.predict(frame, box)
+            gender_str = 'Female' if gender_id == 0 else 'Male'
+
+            # Prepare text and draw on the frame
+            label = f'{gender_str}, {age}'
+            cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
+            cv2.putText(
+                frame,
+                label,
+                (x1, y1 - 10),
+                cv2.FONT_HERSHEY_SIMPLEX,
+                0.8,
+                (0, 255, 0),
+                2,
+            )
+
+        # Display the resulting frame
+        cv2.imshow("Age and Gender Inference (Press 'q' to quit)", frame)
+
+        # Break the loop if 'q' is pressed
+        if cv2.waitKey(1) & 0xFF == ord('q'):
+            break
+
+    # Release resources
+    cap.release()
+    cv2.destroyAllWindows()
+    print('Inference stopped.')

uniface/attribute/base.py

@@ -0,0 +1,92 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from abc import ABC, abstractmethod
+from typing import Any
+
+import numpy as np
+
+
+class Attribute(ABC):
+    """
+    Abstract base class for face attribute models.
+
+    This class defines the common interface that all attribute models
+    (e.g., age-gender, emotion) must implement. It ensures a consistent API
+    across different attribute prediction modules in the library, making them
+    interchangeable and easy to use.
+    """
+
+    @abstractmethod
+    def _initialize_model(self) -> None:
+        """
+        Initializes the underlying model for inference.
+
+        This method should handle loading model weights, creating the
+        inference session (e.g., ONNX Runtime, PyTorch), and any necessary
+        warm-up procedures to prepare the model for prediction.
+        """
+        raise NotImplementedError('Subclasses must implement the _initialize_model method.')
+
+    @abstractmethod
+    def preprocess(self, image: np.ndarray, *args: Any) -> Any:
+        """
+        Preprocesses the input data for the model.
+
+        This method should take a raw image and any other necessary data
+        (like bounding boxes or landmarks) and convert it into the format
+        expected by the model's inference engine (e.g., a blob or tensor).
+
+        Args:
+            image (np.ndarray): The input image containing the face, typically
+                                in BGR format.
+            *args: Additional arguments required for preprocessing, such as
+                   bounding boxes or facial landmarks.
+
+        Returns:
+            The preprocessed data ready for model inference.
+        """
+        raise NotImplementedError('Subclasses must implement the preprocess method.')
+
+    @abstractmethod
+    def postprocess(self, prediction: Any) -> Any:
+        """
+        Postprocesses the raw model output into a human-readable format.
+
+        This method takes the raw output from the model's inference and
+        converts it into a meaningful result, such as an age value, a gender
+        label, or an emotion category.
+
+        Args:
+            prediction (Any): The raw output from the model's inference.
+
+        Returns:
+            The final, processed attributes.
+        """
+        raise NotImplementedError('Subclasses must implement the postprocess method.')
+
+    @abstractmethod
+    def predict(self, image: np.ndarray, *args: Any) -> Any:
+        """
+        Performs end-to-end attribute prediction on a given image.
+
+        This method orchestrates the full pipeline: it calls the preprocess,
+        inference, and postprocess steps to return the final, user-friendly
+        attribute prediction.
+
+        Args:
+            image (np.ndarray): The input image containing the face.
+            *args: Additional data required for prediction, such as a bounding
+                   box or landmarks.
+
+        Returns:
+            The final predicted attributes.
+        """
+        raise NotImplementedError('Subclasses must implement the predict method.')
+
+    def __call__(self, *args, **kwargs) -> Any:
+        """
+        Provides a convenient, callable shortcut for the `predict` method.
+        """
+        return self.predict(*args, **kwargs)

uniface/attribute/emotion.py

@@ -0,0 +1,194 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from typing import List, Tuple, Union
+
+import cv2
+import numpy as np
+import torch
+
+from uniface.attribute.base import Attribute
+from uniface.constants import DDAMFNWeights
+from uniface.face_utils import face_alignment
+from uniface.log import Logger
+from uniface.model_store import verify_model_weights
+
+__all__ = ['Emotion']
+
+
+class Emotion(Attribute):
+    """
+    Emotion recognition model using a TorchScript model.
+
+    This class inherits from the base `Attribute` class and implements the
+    functionality for predicting one of several emotion categories from a face
+    image. It requires 5-point facial landmarks for alignment.
+    """
+
+    def __init__(
+        self,
+        model_weights: DDAMFNWeights = DDAMFNWeights.AFFECNET7,
+        input_size: Tuple[int, int] = (112, 112),
+    ) -> None:
+        """
+        Initializes the emotion recognition model.
+
+        Args:
+            model_weights (DDAMFNWeights): The enum for the model weights to load.
+            input_size (Tuple[int, int]): The expected input size for the model.
+        """
+        Logger.info(f'Initializing Emotion with model={model_weights.name}')
+
+        if torch.backends.mps.is_available():
+            self.device = torch.device('mps')
+        elif torch.cuda.is_available():
+            self.device = torch.device('cuda')
+        else:
+            self.device = torch.device('cpu')
+
+        self.input_size = input_size
+        self.model_path = verify_model_weights(model_weights)
+
+        # Define emotion labels based on the selected model
+        self.emotion_labels = [
+            'Neutral',
+            'Happy',
+            'Sad',
+            'Surprise',
+            'Fear',
+            'Disgust',
+            'Angry',
+        ]
+        if model_weights == DDAMFNWeights.AFFECNET8:
+            self.emotion_labels.append('Contempt')
+
+        self._initialize_model()
+
+    def _initialize_model(self) -> None:
+        """
+        Loads and initializes the TorchScript model for inference.
+        """
+        try:
+            self.model = torch.jit.load(self.model_path, map_location=self.device)
+            self.model.eval()
+            # Warm-up with a dummy input for faster first inference
+            dummy_input = torch.randn(1, 3, *self.input_size).to(self.device)
+            with torch.no_grad():
+                self.model(dummy_input)
+            Logger.info(f'Successfully initialized Emotion model on {self.device}')
+        except Exception as e:
+            Logger.error(f"Failed to load Emotion model from '{self.model_path}'", exc_info=True)
+            raise RuntimeError(f'Failed to initialize Emotion model: {e}') from e
+
+    def preprocess(self, image: np.ndarray, landmark: Union[List, np.ndarray]) -> torch.Tensor:
+        """
+        Aligns the face using landmarks and preprocesses it into a tensor.
+
+        Args:
+            image (np.ndarray): The full input image in BGR format.
+            landmark (Union[List, np.ndarray]): The 5-point facial landmarks.
+
+        Returns:
+            torch.Tensor: The preprocessed image tensor ready for inference.
+        """
+        landmark = np.asarray(landmark)
+
+        aligned_image, _ = face_alignment(image, landmark)
+
+        # Convert BGR to RGB, resize, normalize, and convert to a CHW tensor
+        rgb_image = cv2.cvtColor(aligned_image, cv2.COLOR_BGR2RGB)
+        resized_image = cv2.resize(rgb_image, self.input_size).astype(np.float32) / 255.0
+        mean = np.array([0.485, 0.456, 0.406], dtype=np.float32)
+        std = np.array([0.229, 0.224, 0.225], dtype=np.float32)
+        normalized_image = (resized_image - mean) / std
+        transposed_image = normalized_image.transpose((2, 0, 1))
+
+        return torch.from_numpy(transposed_image).unsqueeze(0).to(self.device)
+
+    def postprocess(self, prediction: torch.Tensor) -> Tuple[str, float]:
+        """
+        Processes the raw model output to get the emotion label and confidence score.
+        """
+        probabilities = torch.nn.functional.softmax(prediction, dim=1).squeeze().cpu().numpy()
+        pred_index = np.argmax(probabilities)
+        emotion_label = self.emotion_labels[pred_index]
+        confidence = float(probabilities[pred_index])
+        return emotion_label, confidence
+
+    def predict(self, image: np.ndarray, landmark: Union[List, np.ndarray]) -> Tuple[str, float]:
+        """
+        Predicts the emotion from a single face specified by its landmarks.
+        """
+        input_tensor = self.preprocess(image, landmark)
+        with torch.no_grad():
+            output = self.model(input_tensor)
+            if isinstance(output, tuple):
+                output = output[0]
+
+        return self.postprocess(output)
+
+
+# TODO: below is only for testing, remove it later
+if __name__ == '__main__':
+    from uniface.constants import RetinaFaceWeights
+    from uniface.detection import create_detector
+
+    print('Initializing models for live inference...')
+    # 1. Initialize the face detector
+    # Using a smaller model for faster real-time performance
+    detector = create_detector(model_name=RetinaFaceWeights.MNET_V2)
+
+    # 2. Initialize the attribute predictor
+    emotion_predictor = Emotion()
+
+    # 3. Start webcam capture
+    cap = cv2.VideoCapture(0)
+    if not cap.isOpened():
+        print('Error: Could not open webcam.')
+        exit()
+
+    print("Starting webcam feed. Press 'q' to quit.")
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            print('Error: Failed to capture frame.')
+            break
+
+        # Detect faces in the current frame.
+        # This method returns a list of dictionaries for each detected face.
+        detections = detector.detect(frame)
+
+        # For each detected face, predict the emotion
+        for detection in detections:
+            box = detection['bbox']
+            landmark = detection['landmarks']
+            x1, y1, x2, y2 = map(int, box)
+
+            # Predict attributes using the landmark
+            emotion, confidence = emotion_predictor.predict(frame, landmark)
+
+            # Prepare text and draw on the frame
+            label = f'{emotion} ({confidence:.2f})'
+            cv2.rectangle(frame, (x1, y1), (x2, y2), (255, 0, 0), 2)
+            cv2.putText(
+                frame,
+                label,
+                (x1, y1 - 10),
+                cv2.FONT_HERSHEY_SIMPLEX,
+                0.8,
+                (255, 0, 0),
+                2,
+            )
+
+        # Display the resulting frame
+        cv2.imshow("Emotion Inference (Press 'q' to quit)", frame)
+
+        # Break the loop if 'q' is pressed
+        if cv2.waitKey(1) & 0xFF == ord('q'):
+            break
+
+    # Release resources
+    cap.release()
+    cv2.destroyAllWindows()
+    print('Inference stopped.')

uniface/common.py

@@ -1,14 +1,23 @@
-# Copyright 2024 Yakhyokhuja Valikhujaev
+# Copyright 2025 Yakhyokhuja Valikhujaev
 # Author: Yakhyokhuja Valikhujaev
 # GitHub: https://github.com/yakhyo
 
-import cv2
-import math
 import itertools
+import math
+from typing import List, Optional, Tuple
+
+import cv2
 import numpy as np
 
-import torch
-from typing import Tuple, List
+__all__ = [
+    'resize_image',
+    'generate_anchors',
+    'non_max_suppression',
+    'decode_boxes',
+    'decode_landmarks',
+    'distance2bbox',
+    'distance2kps',
+]
 
 
 def resize_image(frame, target_shape: Tuple[int, int] = (640, 640)) -> Tuple[np.ndarray, float]:
@@ -44,28 +53,21 @@ def resize_image(frame, target_shape: Tuple[int, int] = (640, 640)) -> Tuple[np.
     return image, resize_factor
 
 
-def generate_anchors(image_size: Tuple[int, int] = (640, 640)) -> torch.Tensor:
+def generate_anchors(image_size: Tuple[int, int] = (640, 640)) -> np.ndarray:
     """
-    Generate anchor boxes for a given image size.
+    Generate anchor boxes for a given image size (RetinaFace specific).
 
     Args:
         image_size (Tuple[int, int]): Input image size (width, height). Defaults to (640, 640).
 
     Returns:
-        torch.Tensor: Anchor box coordinates as a tensor.
+        np.ndarray: Anchor box coordinates as a NumPy array with shape (num_anchors, 4).
     """
-    image_size = image_size
-
     steps = [8, 16, 32]
     min_sizes = [[16, 32], [64, 128], [256, 512]]
 
     anchors = []
-    feature_maps = [
-        [
-            math.ceil(image_size[0] / step),
-            math.ceil(image_size[1] / step)
-        ] for step in steps
-    ]
+    feature_maps = [[math.ceil(image_size[0] / step), math.ceil(image_size[1] / step)] for step in steps]
 
     for k, (map_height, map_width) in enumerate(feature_maps):
         step = steps[k]
@@ -79,20 +81,20 @@ def generate_anchors(image_size: Tuple[int, int] = (640, 640)) -> torch.Tensor:
                 for cy, cx in itertools.product(dense_cy, dense_cx):
                     anchors += [cx, cy, s_kx, s_ky]
 
-    output = torch.Tensor(anchors).view(-1, 4)
+    output = np.array(anchors, dtype=np.float32).reshape(-1, 4)
     return output
 
 
-def nms(dets: List[np.ndarray], threshold: float):
+def non_max_suppression(dets: np.ndarray, threshold: float) -> List[int]:
     """
     Apply Non-Maximum Suppression (NMS) to reduce overlapping bounding boxes based on a threshold.
 
     Args:
-        dets (numpy.ndarray): Array of detections with each row as [x1, y1, x2, y2, score].
+        dets (np.ndarray): Array of detections with each row as [x1, y1, x2, y2, score].
         threshold (float): IoU threshold for suppression.
 
     Returns:
-        list: Indices of bounding boxes retained after suppression.
+        List[int]: Indices of bounding boxes retained after suppression.
     """
     x1 = dets[:, 0]
     y1 = dets[:, 1]
@@ -123,56 +125,119 @@ def nms(dets: List[np.ndarray], threshold: float):
     return keep
 
 
-def decode_boxes(loc, priors, variances=[0.1, 0.2]) -> torch.Tensor:
+def decode_boxes(loc: np.ndarray, priors: np.ndarray, variances: Optional[List[float]] = None) -> np.ndarray:
     """
     Decode locations from predictions using priors to undo
-    the encoding done for offset regression at train time.
+    the encoding done for offset regression at train time (RetinaFace specific).
 
     Args:
-        loc (tensor): Location predictions for loc layers, shape: [num_priors, 4]
-        priors (tensor): Prior boxes in center-offset form, shape: [num_priors, 4]
-        variances (list[float]): Variances of prior boxes
+        loc (np.ndarray): Location predictions for loc layers, shape: [num_priors, 4]
+        priors (np.ndarray): Prior boxes in center-offset form, shape: [num_priors, 4]
+        variances (Optional[List[float]]): Variances of prior boxes. Defaults to [0.1, 0.2].
 
     Returns:
-        tensor: Decoded bounding box predictions
+        np.ndarray: Decoded bounding box predictions with shape [num_priors, 4]
     """
+    if variances is None:
+        variances = [0.1, 0.2]
     # Compute centers of predicted boxes
     cxcy = priors[:, :2] + loc[:, :2] * variances[0] * priors[:, 2:]
 
     # Compute widths and heights of predicted boxes
-    wh = priors[:, 2:] * torch.exp(loc[:, 2:] * variances[1])
+    wh = priors[:, 2:] * np.exp(loc[:, 2:] * variances[1])
 
     # Convert center, size to corner coordinates
-    boxes = torch.empty_like(loc)
+    boxes = np.zeros_like(loc)
     boxes[:, :2] = cxcy - wh / 2  # xmin, ymin
     boxes[:, 2:] = cxcy + wh / 2  # xmax, ymax
 
     return boxes
 
 
-def decode_landmarks(predictions, priors, variances=[0.1, 0.2]) -> torch.Tensor:
+def decode_landmarks(
+    predictions: np.ndarray, priors: np.ndarray, variances: Optional[List[float]] = None
+) -> np.ndarray:
     """
-    Decode landmarks from predictions using prior boxes to reverse the encoding done during training.
+    Decode landmark predictions using prior boxes (RetinaFace specific).
 
     Args:
-        predictions (tensor): Landmark predictions for localization layers.
-            Shape: [num_priors, 10] where each prior contains 5 landmark (x, y) pairs.
-        priors (tensor): Prior boxes in center-offset form.
-            Shape: [num_priors, 4], where each prior has (cx, cy, width, height).
-        variances (list[float]): Variances of the prior boxes to scale the decoded values.
+        predictions (np.ndarray): Landmark predictions, shape: [num_priors, 10]
+        priors (np.ndarray): Prior boxes, shape: [num_priors, 4]
+        variances (Optional[List[float]]): Scaling factors for landmark offsets. Defaults to [0.1, 0.2].
 
     Returns:
-        landmarks (tensor): Decoded landmark predictions.
-            Shape: [num_priors, 10] where each row contains the decoded (x, y) pairs for 5 landmarks.
+        np.ndarray: Decoded landmarks, shape: [num_priors, 10]
     """
+    if variances is None:
+        variances = [0.1, 0.2]
 
-    # Reshape predictions to [num_priors, 5, 2] to handle each pair (x, y) in a batch
-    predictions = predictions.view(predictions.size(0), 5, 2)
+    # Reshape predictions to [num_priors, 5, 2] to process landmark points
+    predictions = predictions.reshape(predictions.shape[0], 5, 2)
 
-    # Perform the same operation on all landmark pairs at once
-    landmarks = priors[:, :2].unsqueeze(1) + predictions * variances[0] * priors[:, 2:].unsqueeze(1)
+    # Expand priors to match (num_priors, 5, 2)
+    priors_xy = np.repeat(priors[:, :2][:, np.newaxis, :], 5, axis=1)  # (num_priors, 5, 2)
+    priors_wh = np.repeat(priors[:, 2:][:, np.newaxis, :], 5, axis=1)  # (num_priors, 5, 2)
+
+    # Compute absolute landmark positions
+    landmarks = priors_xy + predictions * variances[0] * priors_wh
 
     # Flatten back to [num_priors, 10]
-    landmarks = landmarks.view(landmarks.size(0), -1)
+    landmarks = landmarks.reshape(landmarks.shape[0], -1)
 
     return landmarks
+
+
+def distance2bbox(points: np.ndarray, distance: np.ndarray, max_shape: Optional[Tuple[int, int]] = None) -> np.ndarray:
+    """
+    Decode distance prediction to bounding box (SCRFD specific).
+
+    Args:
+        points (np.ndarray): Anchor points with shape (n, 2), [x, y].
+        distance (np.ndarray): Distance from the given point to 4
+            boundaries (left, top, right, bottom) with shape (n, 4).
+        max_shape (Optional[Tuple[int, int]]): Shape of the image (height, width) for clipping.
+
+    Returns:
+        np.ndarray: Decoded bounding boxes with shape (n, 4) as [x1, y1, x2, y2].
+    """
+    x1 = points[:, 0] - distance[:, 0]
+    y1 = points[:, 1] - distance[:, 1]
+    x2 = points[:, 0] + distance[:, 2]
+    y2 = points[:, 1] + distance[:, 3]
+
+    if max_shape is not None:
+        x1 = np.clip(x1, 0, max_shape[1])
+        y1 = np.clip(y1, 0, max_shape[0])
+        x2 = np.clip(x2, 0, max_shape[1])
+        y2 = np.clip(y2, 0, max_shape[0])
+    else:
+        x1 = np.maximum(x1, 0)
+        y1 = np.maximum(y1, 0)
+        x2 = np.maximum(x2, 0)
+        y2 = np.maximum(y2, 0)
+
+    return np.stack([x1, y1, x2, y2], axis=-1)
+
+
+def distance2kps(points: np.ndarray, distance: np.ndarray, max_shape: Optional[Tuple[int, int]] = None) -> np.ndarray:
+    """
+    Decode distance prediction to keypoints (SCRFD specific).
+
+    Args:
+        points (np.ndarray): Anchor points with shape (n, 2), [x, y].
+        distance (np.ndarray): Distance from the given point to keypoints with shape (n, 2k).
+        max_shape (Optional[Tuple[int, int]]): Shape of the image (height, width) for clipping.
+
+    Returns:
+        np.ndarray: Decoded keypoints with shape (n, 2k).
+    """
+    preds = []
+    for i in range(0, distance.shape[1], 2):
+        px = points[:, i % 2] + distance[:, i]
+        py = points[:, i % 2 + 1] + distance[:, i + 1]
+        if max_shape is not None:
+            px = np.clip(px, 0, max_shape[1])
+            py = np.clip(py, 0, max_shape[0])
+        preds.append(px)
+        preds.append(py)
+    return np.stack(preds, axis=-1)

uniface/constants.py

@@ -1,26 +1,210 @@
-# Copyright 2024 Yakhyokhuja Valikhujaev
+# Copyright 2025 Yakhyokhuja Valikhujaev
 # Author: Yakhyokhuja Valikhujaev
 # GitHub: https://github.com/yakhyo
 
+from enum import Enum
 from typing import Dict
 
 
-MODEL_URLS: Dict[str, str] = {
-    'retinaface_mnet025': 'https://github.com/yakhyo/uniface/releases/download/v0.1.1/retinaface_mv1_0.25.onnx',
-    'retinaface_mnet050': 'https://github.com/yakhyo/uniface/releases/download/v0.1.1/retinaface_mv1_0.50.onnx',
-    'retinaface_mnet_v1': 'https://github.com/yakhyo/uniface/releases/download/v0.1.1/retinaface_mv1.onnx',
-    'retinaface_mnet_v2': 'https://github.com/yakhyo/uniface/releases/download/v0.1.1/retinaface_mv2.onnx',
-    'retinaface_r18': 'https://github.com/yakhyo/uniface/releases/download/v0.1.1/retinaface_r18.onnx',
-    'retinaface_r34': 'https://github.com/yakhyo/uniface/releases/download/v0.1.1/retinaface_r34.onnx'
+# fmt: off
+class SphereFaceWeights(str, Enum):
+    """
+    Trained on MS1M V2 dataset with 5.8 million images of 85k identities.
+    https://github.com/yakhyo/face-recognition
+    """
+    SPHERE20      = "sphere20"
+    SPHERE36      = "sphere36"
+
+class MobileFaceWeights(str, Enum):
+    """
+    Trained on MS1M V2 dataset with 5.8 million images of 85k identities.
+    https://github.com/yakhyo/face-recognition
+    """
+    MNET_025      = "mobilenetv1_025"
+    MNET_V2       = "mobilenetv2"
+    MNET_V3_SMALL = "mobilenetv3_small"
+    MNET_V3_LARGE = "mobilenetv3_large"
+
+class ArcFaceWeights(str, Enum):
+    """
+    Pretrained weights from ArcFace model (insightface).
+    https://github.com/deepinsight/insightface
+    """
+    MNET   = "arcface_mnet"
+    RESNET = "arcface_resnet"
+
+class RetinaFaceWeights(str, Enum):
+    """
+    Trained on WIDER FACE dataset.
+    https://github.com/yakhyo/retinaface-pytorch
+    """
+    MNET_025 =  "retinaface_mnet025"
+    MNET_050 =  "retinaface_mnet050"
+    MNET_V1  =  "retinaface_mnet_v1"
+    MNET_V2  =  "retinaface_mnet_v2"
+    RESNET18 =  "retinaface_r18"
+    RESNET34 =  "retinaface_r34"
+
+
+class SCRFDWeights(str, Enum):
+    """
+    Trained on WIDER FACE dataset.
+    https://github.com/deepinsight/insightface
+    """
+    SCRFD_10G_KPS  = "scrfd_10g"
+    SCRFD_500M_KPS = "scrfd_500m"
+
+
+class YOLOv5FaceWeights(str, Enum):
+    """
+    Trained on WIDER FACE dataset.
+    Original implementation: https://github.com/deepcam-cn/yolov5-face
+    Exported to ONNX from: https://github.com/yakhyo/yolov5-face-onnx-inference
+
+    Model Performance (WIDER FACE):
+    - YOLOV5N: 11MB, 93.61% Easy / 91.52% Medium / 80.53% Hard
+    - YOLOV5S: 28MB, 94.33% Easy / 92.61% Medium / 83.15% Hard
+    - YOLOV5M: 82MB, 95.30% Easy / 93.76% Medium / 85.28% Hard
+    """
+    YOLOV5N = "yolov5n"
+    YOLOV5S = "yolov5s"
+    YOLOV5M = "yolov5m"
+
+
+class DDAMFNWeights(str, Enum):
+    """
+    Trained on AffectNet dataset.
+    https://github.com/SainingZhang/DDAMFN/tree/main/DDAMFN
+    """
+    AFFECNET7 = "affecnet7"
+    AFFECNET8 = "affecnet8"
+
+
+class AgeGenderWeights(str, Enum):
+    """
+    Trained on CelebA dataset.
+    https://github.com/deepinsight/insightface
+    """
+    DEFAULT = "age_gender"
+
+
+class LandmarkWeights(str, Enum):
+    """
+    MobileNet 0.5 from Insightface
+    https://github.com/deepinsight/insightface/tree/master/alignment/coordinate_reg
+    """
+    DEFAULT = "2d_106"
+
+
+class GazeWeights(str, Enum):
+    """
+    MobileGaze: Real-Time Gaze Estimation models.
+    Trained on Gaze360 dataset.
+    https://github.com/yakhyo/gaze-estimation
+    """
+    RESNET18     = "gaze_resnet18"
+    RESNET34     = "gaze_resnet34"
+    RESNET50     = "gaze_resnet50"
+    MOBILENET_V2 = "gaze_mobilenetv2"
+    MOBILEONE_S0 = "gaze_mobileone_s0"
+
+
+class ParsingWeights(str, Enum):
+    """
+    Face Parsing: Semantic Segmentation of Facial Components.
+    Trained on CelebAMask-HQ dataset.
+    https://github.com/yakhyo/face-parsing
+    """
+    RESNET18 = "parsing_resnet18"
+    RESNET34 = "parsing_resnet34"
+
+
+MODEL_URLS: Dict[Enum, str] = {
+    # RetinaFace
+    RetinaFaceWeights.MNET_025:      'https://github.com/yakhyo/uniface/releases/download/weights/retinaface_mv1_0.25.onnx',
+    RetinaFaceWeights.MNET_050:      'https://github.com/yakhyo/uniface/releases/download/weights/retinaface_mv1_0.50.onnx',
+    RetinaFaceWeights.MNET_V1:       'https://github.com/yakhyo/uniface/releases/download/weights/retinaface_mv1.onnx',
+    RetinaFaceWeights.MNET_V2:       'https://github.com/yakhyo/uniface/releases/download/weights/retinaface_mv2.onnx',
+    RetinaFaceWeights.RESNET18:      'https://github.com/yakhyo/uniface/releases/download/weights/retinaface_r18.onnx',
+    RetinaFaceWeights.RESNET34:      'https://github.com/yakhyo/uniface/releases/download/weights/retinaface_r34.onnx',
+    # MobileFace
+    MobileFaceWeights.MNET_025:      'https://github.com/yakhyo/uniface/releases/download/weights/mobilenetv1_0.25.onnx',
+    MobileFaceWeights.MNET_V2:       'https://github.com/yakhyo/uniface/releases/download/weights/mobilenetv2.onnx',
+    MobileFaceWeights.MNET_V3_SMALL: 'https://github.com/yakhyo/uniface/releases/download/weights/mobilenetv3_small.onnx',
+    MobileFaceWeights.MNET_V3_LARGE: 'https://github.com/yakhyo/uniface/releases/download/weights/mobilenetv3_large.onnx',
+    # SphereFace
+    SphereFaceWeights.SPHERE20:      'https://github.com/yakhyo/uniface/releases/download/weights/sphere20.onnx',
+    SphereFaceWeights.SPHERE36:      'https://github.com/yakhyo/uniface/releases/download/weights/sphere36.onnx',
+    # ArcFace
+    ArcFaceWeights.MNET:             'https://github.com/yakhyo/uniface/releases/download/weights/w600k_mbf.onnx',
+    ArcFaceWeights.RESNET:           'https://github.com/yakhyo/uniface/releases/download/weights/w600k_r50.onnx',
+    # SCRFD
+    SCRFDWeights.SCRFD_10G_KPS:      'https://github.com/yakhyo/uniface/releases/download/weights/scrfd_10g_kps.onnx',
+    SCRFDWeights.SCRFD_500M_KPS:     'https://github.com/yakhyo/uniface/releases/download/weights/scrfd_500m_kps.onnx',
+    # YOLOv5-Face
+    YOLOv5FaceWeights.YOLOV5N:       'https://github.com/yakhyo/yolov5-face-onnx-inference/releases/download/weights/yolov5n_face.onnx',
+    YOLOv5FaceWeights.YOLOV5S:       'https://github.com/yakhyo/yolov5-face-onnx-inference/releases/download/weights/yolov5s_face.onnx',
+    YOLOv5FaceWeights.YOLOV5M:       'https://github.com/yakhyo/yolov5-face-onnx-inference/releases/download/weights/yolov5m_face.onnx',
+    # DDAFM
+    DDAMFNWeights.AFFECNET7:         'https://github.com/yakhyo/uniface/releases/download/weights/affecnet7.script',
+    DDAMFNWeights.AFFECNET8:         'https://github.com/yakhyo/uniface/releases/download/weights/affecnet8.script',
+    # AgeGender
+    AgeGenderWeights.DEFAULT:        'https://github.com/yakhyo/uniface/releases/download/weights/genderage.onnx',
+    # Landmarks
+    LandmarkWeights.DEFAULT:         'https://github.com/yakhyo/uniface/releases/download/weights/2d106det.onnx',
+    # Gaze (MobileGaze)
+    GazeWeights.RESNET18:            'https://github.com/yakhyo/gaze-estimation/releases/download/weights/resnet18_gaze.onnx',
+    GazeWeights.RESNET34:            'https://github.com/yakhyo/gaze-estimation/releases/download/weights/resnet34_gaze.onnx',
+    GazeWeights.RESNET50:            'https://github.com/yakhyo/gaze-estimation/releases/download/weights/resnet50_gaze.onnx',
+    GazeWeights.MOBILENET_V2:        'https://github.com/yakhyo/gaze-estimation/releases/download/weights/mobilenetv2_gaze.onnx',
+    GazeWeights.MOBILEONE_S0:        'https://github.com/yakhyo/gaze-estimation/releases/download/weights/mobileone_s0_gaze.onnx',
+    # Parsing
+    ParsingWeights.RESNET18:         'https://github.com/yakhyo/face-parsing/releases/download/weights/resnet18.onnx',
+    ParsingWeights.RESNET34:         'https://github.com/yakhyo/face-parsing/releases/download/weights/resnet34.onnx',
 }
 
-MODEL_SHA256: Dict[str, str] = {
-    'retinaface_mnet025': 'b7a7acab55e104dce6f32cdfff929bd83946da5cd869b9e2e9bdffafd1b7e4a5',
-    'retinaface_mnet050': 'd8977186f6037999af5b4113d42ba77a84a6ab0c996b17c713cc3d53b88bfc37',
-    'retinaface_mnet_v1': '75c961aaf0aff03d13c074e9ec656e5510e174454dd4964a161aab4fe5f04153',
-    'retinaface_mnet_v2': '3ca44c045651cabeed1193a1fae8946ad1f3a55da8fa74b341feab5a8319f757',
-    'retinaface_r18': 'e8b5ddd7d2c3c8f7c942f9f10cec09d8e319f78f09725d3f709631de34fb649d',
-    'retinaface_r34': 'bd0263dc2a465d32859555cb1741f2d98991eb0053696e8ee33fec583d30e630'
+MODEL_SHA256: Dict[Enum, str] = {
+    # RetinaFace
+    RetinaFaceWeights.MNET_025:      'b7a7acab55e104dce6f32cdfff929bd83946da5cd869b9e2e9bdffafd1b7e4a5',
+    RetinaFaceWeights.MNET_050:      'd8977186f6037999af5b4113d42ba77a84a6ab0c996b17c713cc3d53b88bfc37',
+    RetinaFaceWeights.MNET_V1:       '75c961aaf0aff03d13c074e9ec656e5510e174454dd4964a161aab4fe5f04153',
+    RetinaFaceWeights.MNET_V2:       '3ca44c045651cabeed1193a1fae8946ad1f3a55da8fa74b341feab5a8319f757',
+    RetinaFaceWeights.RESNET18:      'e8b5ddd7d2c3c8f7c942f9f10cec09d8e319f78f09725d3f709631de34fb649d',
+    RetinaFaceWeights.RESNET34:      'bd0263dc2a465d32859555cb1741f2d98991eb0053696e8ee33fec583d30e630',
+    # MobileFace
+    MobileFaceWeights.MNET_025:      'eeda7d23d9c2b40cf77fa8da8e895b5697465192648852216074679657f8ee8b',
+    MobileFaceWeights.MNET_V2:       '38b148284dd48cc898d5d4453104252fbdcbacc105fe3f0b80e78954d9d20d89',
+    MobileFaceWeights.MNET_V3_SMALL: 'd4acafa1039a82957aa8a9a1dac278a401c353a749c39df43de0e29cc1c127c3',
+    MobileFaceWeights.MNET_V3_LARGE: '0e48f8e11f070211716d03e5c65a3db35a5e917cfb5bc30552358629775a142a',
+    # SphereFace
+    SphereFaceWeights.SPHERE20:      'c02878cf658eb1861f580b7e7144b0d27cc29c440bcaa6a99d466d2854f14c9d',
+    SphereFaceWeights.SPHERE36:      '13b3890cd5d7dec2b63f7c36fd7ce07403e5a0bbb701d9647c0289e6cbe7bb20',
+    # ArcFace
+    ArcFaceWeights.MNET:             '9cc6e4a75f0e2bf0b1aed94578f144d15175f357bdc05e815e5c4a02b319eb4f',
+    ArcFaceWeights.RESNET:           '4c06341c33c2ca1f86781dab0e829f88ad5b64be9fba56e56bc9ebdefc619e43',
+    # SCRFD
+    SCRFDWeights.SCRFD_10G_KPS:      '5838f7fe053675b1c7a08b633df49e7af5495cee0493c7dcf6697200b85b5b91',
+    SCRFDWeights.SCRFD_500M_KPS:     '5e4447f50245bbd7966bd6c0fa52938c61474a04ec7def48753668a9d8b4ea3a',
+    # YOLOv5-Face
+    YOLOv5FaceWeights.YOLOV5N:       'eb244a06e36999db732b317c2b30fa113cd6cfc1a397eaf738f2d6f33c01f640',
+    YOLOv5FaceWeights.YOLOV5S:       'fc682801cd5880e1e296184a14aea0035486b5146ec1a1389d2e7149cb134bb2',
+    YOLOv5FaceWeights.YOLOV5M:       '04302ce27a15bde3e20945691b688e2dd018a10e92dd8932146bede6a49207b2',
+    # DDAFM
+    DDAMFNWeights.AFFECNET7:         '10535bf8b6afe8e9d6ae26cea6c3add9a93036e9addb6adebfd4a972171d015d',
+    DDAMFNWeights.AFFECNET8:         '8c66963bc71db42796a14dfcbfcd181b268b65a3fc16e87147d6a3a3d7e0f487',
+    # AgeGender
+    AgeGenderWeights.DEFAULT:        '4fde69b1c810857b88c64a335084f1c3fe8f01246c9a191b48c7bb756d6652fb',
+    # Landmark
+    LandmarkWeights.DEFAULT:         'f001b856447c413801ef5c42091ed0cd516fcd21f2d6b79635b1e733a7109dbf',
+    # MobileGaze (trained on Gaze360)
+    GazeWeights.RESNET18:            '23d5d7e4f6f40dce8c35274ce9d08b45b9e22cbaaf5af73182f473229d713d31',
+    GazeWeights.RESNET34:            '4457ee5f7acd1a5ab02da4b61f02fc3a0b17adbf3844dd0ba3cd4288f2b5e1de',
+    GazeWeights.RESNET50:            'e1eaf98f5ec7c89c6abe7cfe39f7be83e747163f98d1ff945c0603b3c521be22',
+    GazeWeights.MOBILENET_V2:        'fdcdb84e3e6421b5a79e8f95139f249fc258d7f387eed5ddac2b80a9a15ce076',
+    GazeWeights.MOBILEONE_S0:        'c0b5a4f4a0ffd24f76ab3c1452354bb2f60110899fd9a88b464c75bafec0fde8',
+    # Face Parsing
+    ParsingWeights.RESNET18:         '0d9bd318e46987c3bdbfacae9e2c0f461cae1c6ac6ea6d43bbe541a91727e33f',
+    ParsingWeights.RESNET34:         '5b805bba7b5660ab7070b5a381dcf75e5b3e04199f1e9387232a77a00095102e',
 }
 
 CHUNK_SIZE = 8192

uniface/detection/__init__.py

@@ -0,0 +1,168 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+
+from typing import Any, Dict, List
+
+import numpy as np
+
+from .base import BaseDetector
+from .retinaface import RetinaFace
+from .scrfd import SCRFD
+from .yolov5 import YOLOv5Face
+
+# Global cache for detector instances
+_detector_cache: Dict[str, BaseDetector] = {}
+
+
+def detect_faces(image: np.ndarray, method: str = 'retinaface', **kwargs) -> List[Dict[str, Any]]:
+    """
+    High-level face detection function.
+
+    Args:
+        image (np.ndarray): Input image as numpy array.
+        method (str): Detection method to use. Options: 'retinaface', 'scrfd', 'yolov5face'.
+        **kwargs: Additional arguments passed to the detector.
+
+    Returns:
+        List[Dict[str, Any]]: A list of dictionaries, where each dictionary represents a detected face and contains:
+            - 'bbox' (List[float]): [x1, y1, x2, y2] bounding box coordinates.
+            - 'confidence' (float): The confidence score of the detection.
+            - 'landmarks' (List[List[float]]): 5-point facial landmarks.
+
+    Example:
+        >>> from uniface import detect_faces
+        >>> image = cv2.imread("your_image.jpg")
+        >>> faces = detect_faces(image, method='retinaface', conf_thresh=0.8)
+        >>> for face in faces:
+        ...     print(f"Found face with confidence: {face['confidence']}")
+        ...     print(f"BBox: {face['bbox']}")
+    """
+    method_name = method.lower()
+
+    sorted_kwargs = sorted(kwargs.items())
+    cache_key = f'{method_name}_{str(sorted_kwargs)}'
+
+    if cache_key not in _detector_cache:
+        # Pass kwargs to create the correctly configured detector
+        _detector_cache[cache_key] = create_detector(method, **kwargs)
+
+    detector = _detector_cache[cache_key]
+    return detector.detect(image)
+
+
+def create_detector(method: str = 'retinaface', **kwargs) -> BaseDetector:
+    """
+    Factory function to create face detectors.
+
+    Args:
+        method (str): Detection method. Options:
+            - 'retinaface': RetinaFace detector (default)
+            - 'scrfd': SCRFD detector (fast and accurate)
+            - 'yolov5face': YOLOv5-Face detector (accurate with landmarks)
+        **kwargs: Detector-specific parameters
+
+    Returns:
+        BaseDetector: Initialized detector instance
+
+    Raises:
+        ValueError: If method is not supported
+
+    Examples:
+        >>> # Basic usage
+        >>> detector = create_detector('retinaface')
+
+        >>> # SCRFD detector with custom parameters
+        >>> detector = create_detector(
+        ...     'scrfd',
+        ...     model_name=SCRFDWeights.SCRFD_10G_KPS,
+        ...     conf_thresh=0.8,
+        ...     input_size=(640, 640)
+        ... )
+
+        >>> # RetinaFace detector
+        >>> detector = create_detector(
+        ...     'retinaface',
+        ...     model_name=RetinaFaceWeights.MNET_V2,
+        ...     conf_thresh=0.8,
+        ...     nms_thresh=0.4
+        ... )
+
+        >>> # YOLOv5-Face detector
+        >>> detector = create_detector(
+        ...     'yolov5face',
+        ...     model_name=YOLOv5FaceWeights.YOLOV5S,
+        ...     conf_thresh=0.25,
+        ...     nms_thresh=0.45
+        ... )
+    """
+    method = method.lower()
+
+    if method == 'retinaface':
+        return RetinaFace(**kwargs)
+
+    elif method == 'scrfd':
+        return SCRFD(**kwargs)
+
+    elif method == 'yolov5face':
+        return YOLOv5Face(**kwargs)
+
+    else:
+        available_methods = ['retinaface', 'scrfd', 'yolov5face']
+        raise ValueError(f"Unsupported detection method: '{method}'. Available methods: {available_methods}")
+
+
+def list_available_detectors() -> Dict[str, Dict[str, Any]]:
+    """
+    List all available detection methods with their descriptions and parameters.
+
+    Returns:
+        Dict[str, Dict[str, Any]]: Dictionary of detector information
+    """
+    return {
+        'retinaface': {
+            'description': 'RetinaFace detector with high accuracy',
+            'supports_landmarks': True,
+            'paper': 'https://arxiv.org/abs/1905.00641',
+            'default_params': {
+                'model_name': 'mnet_v2',
+                'conf_thresh': 0.5,
+                'nms_thresh': 0.4,
+                'input_size': (640, 640),
+            },
+        },
+        'scrfd': {
+            'description': 'SCRFD detector - fast and accurate with efficient architecture',
+            'supports_landmarks': True,
+            'paper': 'https://arxiv.org/abs/2105.04714',
+            'default_params': {
+                'model_name': 'scrfd_10g_kps',
+                'conf_thresh': 0.5,
+                'nms_thresh': 0.4,
+                'input_size': (640, 640),
+            },
+        },
+        'yolov5face': {
+            'description': 'YOLOv5-Face detector - accurate face detection with landmarks',
+            'supports_landmarks': True,
+            'paper': 'https://arxiv.org/abs/2105.12931',
+            'default_params': {
+                'model_name': 'yolov5s_face',
+                'conf_thresh': 0.25,
+                'nms_thresh': 0.45,
+                'input_size': 640,
+            },
+        },
+    }
+
+
+__all__ = [
+    'detect_faces',
+    'create_detector',
+    'list_available_detectors',
+    'SCRFD',
+    'RetinaFace',
+    'YOLOv5Face',
+    'BaseDetector',
+]

uniface/detection/base.py

@@ -0,0 +1,104 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from abc import ABC, abstractmethod
+from typing import Any, Dict, List
+
+import numpy as np
+
+
+class BaseDetector(ABC):
+    """
+    Abstract base class for all face detectors.
+
+    This class defines the interface that all face detectors must implement,
+    ensuring consistency across different detection methods.
+    """
+
+    def __init__(self, **kwargs):
+        """Initialize the detector with configuration parameters."""
+        self.config = kwargs
+
+    @abstractmethod
+    def detect(self, image: np.ndarray, **kwargs) -> List[Dict[str, Any]]:
+        """
+        Detect faces in an image.
+
+        Args:
+            image (np.ndarray): Input image as numpy array with shape (H, W, C)
+            **kwargs: Additional detection parameters
+
+        Returns:
+            List[Dict[str, Any]]: List of detected faces, where each dictionary contains:
+                - 'bbox' (np.ndarray): Bounding box coordinates with shape (4,) as [x1, y1, x2, y2]
+                - 'confidence' (float): Detection confidence score (0.0 to 1.0)
+                - 'landmarks' (np.ndarray): Facial landmarks with shape (5, 2) for 5-point landmarks
+                  or (68, 2) for 68-point landmarks. Empty array if not supported.
+
+        Example:
+            >>> faces = detector.detect(image)
+            >>> for face in faces:
+            ...     bbox = face['bbox']  # np.ndarray with shape (4,)
+            ...     confidence = face['confidence']  # float
+            ...     landmarks = face['landmarks']  # np.ndarray with shape (5, 2)
+        """
+        pass
+
+    @abstractmethod
+    def preprocess(self, image: np.ndarray) -> np.ndarray:
+        """
+        Preprocess input image for detection.
+
+        Args:
+            image (np.ndarray): Input image
+
+        Returns:
+            np.ndarray: Preprocessed image tensor
+        """
+        pass
+
+    @abstractmethod
+    def postprocess(self, outputs, **kwargs) -> Any:
+        """
+        Postprocess model outputs to get final detections.
+
+        Args:
+            outputs: Raw model outputs
+            **kwargs: Additional postprocessing parameters
+
+        Returns:
+            Any: Processed outputs (implementation-specific format, typically tuple of arrays)
+        """
+        pass
+
+    def __str__(self) -> str:
+        """String representation of the detector."""
+        return f'{self.__class__.__name__}({self.config})'
+
+    def __repr__(self) -> str:
+        """Detailed string representation."""
+        return self.__str__()
+
+    @property
+    def supports_landmarks(self) -> bool:
+        """
+        Whether this detector supports landmark detection.
+
+        Returns:
+            bool: True if landmarks are supported, False otherwise
+        """
+        return hasattr(self, '_supports_landmarks') and self._supports_landmarks
+
+    def get_info(self) -> Dict[str, Any]:
+        """
+        Get detector information and configuration.
+
+        Returns:
+            Dict[str, Any]: Detector information
+        """
+        return {
+            'name': self.__class__.__name__,
+            'supports_landmarks': self._supports_landmarks,
+            'config': self.config,
+        }

uniface/detection/retinaface.py

@@ -0,0 +1,383 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from typing import Any, Dict, List, Literal, Tuple
+
+import numpy as np
+
+from uniface.common import (
+    decode_boxes,
+    decode_landmarks,
+    generate_anchors,
+    non_max_suppression,
+    resize_image,
+)
+from uniface.constants import RetinaFaceWeights
+from uniface.log import Logger
+from uniface.model_store import verify_model_weights
+from uniface.onnx_utils import create_onnx_session
+
+from .base import BaseDetector
+
+
+class RetinaFace(BaseDetector):
+    """
+    Face detector based on the RetinaFace architecture.
+
+    Title: "RetinaFace: Single-stage Dense Face Localisation in the Wild"
+    Paper: https://arxiv.org/abs/1905.00641
+    Code: https://github.com/yakhyo/retinaface-pytorch
+
+    Args:
+        model_name (RetinaFaceWeights): Model weights to use. Defaults to `RetinaFaceWeights.MNET_V2`.
+        conf_thresh (float): Confidence threshold for filtering detections. Defaults to 0.5.
+        nms_thresh (float): Non-maximum suppression (NMS) IoU threshold. Defaults to 0.4.
+        input_size (Tuple[int, int]): Fixed input size (width, height) if `dynamic_size=False`.
+            Defaults to (640, 640).
+            Note: Non-default sizes may cause slower inference and CoreML compatibility issues.
+        **kwargs: Advanced options:
+            pre_nms_topk (int): Number of top-scoring boxes considered before NMS. Defaults to 5000.
+            post_nms_topk (int): Max number of detections kept after NMS. Defaults to 750.
+            dynamic_size (bool): If True, generate anchors dynamically per input image. Defaults to False.
+
+    Attributes:
+        model_name (RetinaFaceWeights): Selected model variant.
+        conf_thresh (float): Threshold for confidence-based filtering.
+        nms_thresh (float): IoU threshold used for NMS.
+        pre_nms_topk (int): Limit on proposals before applying NMS.
+        post_nms_topk (int): Limit on retained detections after NMS.
+        dynamic_size (bool): Flag indicating dynamic or static input sizing.
+        input_size (Tuple[int, int]): Static input size if `dynamic_size=False`.
+        _model_path (str): Absolute path to the verified model weights.
+        _priors (np.ndarray): Precomputed anchor boxes (if static size).
+        _supports_landmarks (bool): Indicates landmark prediction support.
+
+    Raises:
+        ValueError: If the model weights are invalid or not found.
+        RuntimeError: If the ONNX model fails to load or initialize.
+    """
+
+    def __init__(
+        self,
+        *,
+        model_name: RetinaFaceWeights = RetinaFaceWeights.MNET_V2,
+        conf_thresh: float = 0.5,
+        nms_thresh: float = 0.4,
+        input_size: Tuple[int, int] = (640, 640),
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(
+            model_name=model_name,
+            conf_thresh=conf_thresh,
+            nms_thresh=nms_thresh,
+            input_size=input_size,
+            **kwargs,
+        )
+        self._supports_landmarks = True  # RetinaFace supports landmarks
+
+        self.model_name = model_name
+        self.conf_thresh = conf_thresh
+        self.nms_thresh = nms_thresh
+        self.input_size = input_size
+
+        # Advanced options from kwargs
+        self.pre_nms_topk = kwargs.get('pre_nms_topk', 5000)
+        self.post_nms_topk = kwargs.get('post_nms_topk', 750)
+        self.dynamic_size = kwargs.get('dynamic_size', False)
+
+        Logger.info(
+            f'Initializing RetinaFace with model={self.model_name}, conf_thresh={self.conf_thresh}, '
+            f'nms_thresh={self.nms_thresh}, input_size={self.input_size}'
+        )
+
+        # Get path to model weights
+        self._model_path = verify_model_weights(self.model_name)
+        Logger.info(f'Verified model weights located at: {self._model_path}')
+
+        # Precompute anchors if using static size
+        if not self.dynamic_size and self.input_size is not None:
+            self._priors = generate_anchors(image_size=self.input_size)
+            Logger.debug('Generated anchors for static input size.')
+
+        # Initialize model
+        self._initialize_model(self._model_path)
+
+    def _initialize_model(self, model_path: str) -> None:
+        """
+        Initializes an ONNX model session from the given path.
+
+        Args:
+            model_path (str): The file path to the ONNX model.
+
+        Raises:
+            RuntimeError: If the model fails to load, logs an error and raises an exception.
+        """
+        try:
+            self.session = create_onnx_session(model_path)
+            self.input_names = self.session.get_inputs()[0].name
+            self.output_names = [x.name for x in self.session.get_outputs()]
+            Logger.info(f'Successfully initialized the model from {model_path}')
+        except Exception as e:
+            Logger.error(f"Failed to load model from '{model_path}': {e}", exc_info=True)
+            raise RuntimeError(f"Failed to initialize model session for '{model_path}'") from e
+
+    def preprocess(self, image: np.ndarray) -> np.ndarray:
+        """Preprocess input image for model inference.
+
+        Args:
+            image (np.ndarray): Input image.
+
+        Returns:
+            np.ndarray: Preprocessed image tensor with shape (1, C, H, W)
+        """
+        image = np.float32(image) - np.array([104, 117, 123], dtype=np.float32)
+        image = image.transpose(2, 0, 1)  # HWC to CHW
+        image = np.expand_dims(image, axis=0)  # Add batch dimension (1, C, H, W)
+        return image
+
+    def inference(self, input_tensor: np.ndarray) -> List[np.ndarray]:
+        """Perform model inference on the preprocessed image tensor.
+
+        Args:
+            input_tensor (np.ndarray): Preprocessed input tensor.
+
+        Returns:
+            Tuple[np.ndarray, np.ndarray]: Raw model outputs.
+        """
+        return self.session.run(self.output_names, {self.input_names: input_tensor})
+
+    def detect(
+        self,
+        image: np.ndarray,
+        *,
+        max_num: int = 0,
+        metric: Literal['default', 'max'] = 'max',
+        center_weight: float = 2.0,
+    ) -> List[Dict[str, Any]]:
+        """
+        Perform face detection on an input image and return bounding boxes and facial landmarks.
+
+        Args:
+            image (np.ndarray): Input image as a NumPy array of shape (H, W, C).
+            max_num (int): Maximum number of detections to return. Use 0 to return all detections. Defaults to 0.
+            metric (Literal["default", "max"]): Metric for ranking detections when `max_num` is limited.
+                - "default": Prioritize detections closer to the image center.
+                - "max": Prioritize detections with larger bounding box areas.
+            center_weight (float): Weight for penalizing detections farther from the image center
+                when using the "default" metric. Defaults to 2.0.
+
+        Returns:
+            List[Dict[str, Any]]: List of face detection dictionaries, each containing:
+                - 'bbox' (np.ndarray): Bounding box coordinates with shape (4,) as [x1, y1, x2, y2]
+                - 'confidence' (float): Detection confidence score (0.0 to 1.0)
+                - 'landmarks' (np.ndarray): 5-point facial landmarks with shape (5, 2)
+
+        Example:
+            >>> faces = detector.detect(image)
+            >>> for face in faces:
+            ...     bbox = face['bbox']  # np.ndarray with shape (4,)
+            ...     confidence = face['confidence']  # float
+            ...     landmarks = face['landmarks']  # np.ndarray with shape (5, 2)
+            ...     # Can pass landmarks directly to recognition
+            ...     embedding = recognizer.get_normalized_embedding(image, landmarks)
+        """
+
+        original_height, original_width = image.shape[:2]
+
+        if self.dynamic_size:
+            height, width, _ = image.shape
+            self._priors = generate_anchors(image_size=(height, width))  # generate anchors for each input image
+            resize_factor = 1.0  # No resizing
+        else:
+            image, resize_factor = resize_image(image, target_shape=self.input_size)
+
+        height, width, _ = image.shape
+        image_tensor = self.preprocess(image)
+
+        # ONNXRuntime inference
+        outputs = self.inference(image_tensor)
+
+        # Postprocessing
+        detections, landmarks = self.postprocess(outputs, resize_factor, shape=(width, height))
+
+        if max_num > 0 and detections.shape[0] > max_num:
+            # Calculate area of detections
+            areas = (detections[:, 2] - detections[:, 0]) * (detections[:, 3] - detections[:, 1])
+
+            # Calculate offsets from image center
+            center = (original_height // 2, original_width // 2)
+            offsets = np.vstack(
+                [
+                    (detections[:, 0] + detections[:, 2]) / 2 - center[1],
+                    (detections[:, 1] + detections[:, 3]) / 2 - center[0],
+                ]
+            )
+            offset_dist_squared = np.sum(np.power(offsets, 2.0), axis=0)
+
+            # Calculate scores based on the chosen metric
+            if metric == 'max':
+                scores = areas
+            else:
+                scores = areas - offset_dist_squared * center_weight
+
+            # Sort by scores and select top `max_num`
+            sorted_indices = np.argsort(scores)[::-1][:max_num]
+
+            detections = detections[sorted_indices]
+            landmarks = landmarks[sorted_indices]
+
+        faces = []
+        for i in range(detections.shape[0]):
+            face_dict = {
+                'bbox': detections[i, :4],
+                'confidence': float(detections[i, 4]),
+                'landmarks': landmarks[i],
+            }
+            faces.append(face_dict)
+
+        return faces
+
+    def postprocess(
+        self, outputs: List[np.ndarray], resize_factor: float, shape: Tuple[int, int]
+    ) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        Process the model outputs into final detection results.
+
+        Args:
+            outputs (List[np.ndarray]): Raw outputs from the detection model.
+                - outputs[0]: Location predictions (bounding box coordinates).
+                - outputs[1]: Class confidence scores.
+                - outputs[2]: Landmark predictions.
+            resize_factor (float): Factor used to resize the input image during preprocessing.
+            shape (Tuple[int, int]): Original shape of the image as (height, width).
+
+        Returns:
+            Tuple[np.ndarray, np.ndarray]: Processed results containing:
+                - detections (np.ndarray): Array of detected bounding boxes with confidence scores.
+                Shape: (num_detections, 5), where each row is [x_min, y_min, x_max, y_max, score].
+                - landmarks (np.ndarray): Array of detected facial landmarks.
+                Shape: (num_detections, 5, 2), where each row contains 5 landmark points (x, y).
+        """
+        loc, conf, landmarks = (
+            outputs[0].squeeze(0),
+            outputs[1].squeeze(0),
+            outputs[2].squeeze(0),
+        )
+
+        # Decode boxes and landmarks
+        boxes = decode_boxes(loc, self._priors)
+        landmarks = decode_landmarks(landmarks, self._priors)
+
+        boxes, landmarks = self._scale_detections(boxes, landmarks, resize_factor, shape=(shape[0], shape[1]))
+
+        # Extract confidence scores for the face class
+        scores = conf[:, 1]
+        mask = scores > self.conf_thresh
+
+        # Filter by confidence threshold
+        boxes, landmarks, scores = boxes[mask], landmarks[mask], scores[mask]
+
+        # Sort by scores
+        order = scores.argsort()[::-1][: self.pre_nms_topk]
+        boxes, landmarks, scores = boxes[order], landmarks[order], scores[order]
+
+        # Apply NMS
+        detections = np.hstack((boxes, scores[:, np.newaxis])).astype(np.float32, copy=False)
+        keep = non_max_suppression(detections, self.nms_thresh)
+        detections, landmarks = detections[keep], landmarks[keep]
+
+        # Keep top-k detections
+        detections, landmarks = (
+            detections[: self.post_nms_topk],
+            landmarks[: self.post_nms_topk],
+        )
+
+        landmarks = landmarks.reshape(-1, 5, 2).astype(np.float32)
+
+        return detections, landmarks
+
+    def _scale_detections(
+        self,
+        boxes: np.ndarray,
+        landmarks: np.ndarray,
+        resize_factor: float,
+        shape: Tuple[int, int],
+    ) -> Tuple[np.ndarray, np.ndarray]:
+        # Scale bounding boxes and landmarks to the original image size.
+        bbox_scale = np.array([shape[0], shape[1]] * 2)
+        boxes = boxes * bbox_scale / resize_factor
+
+        landmark_scale = np.array([shape[0], shape[1]] * 5)
+        landmarks = landmarks * landmark_scale / resize_factor
+
+        return boxes, landmarks
+
+
+# TODO: below is only for testing, remove it later
+def draw_bbox(frame, bbox, score, color=(0, 255, 0), thickness=2):
+    x1, y1, x2, y2 = map(int, bbox)  # Unpack 4 bbox values
+    cv2.rectangle(frame, (x1, y1), (x2, y2), color, thickness)
+    cv2.putText(frame, f'{score:.2f}', (x1, y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, color, 1)
+
+
+def draw_keypoints(frame, points, color=(0, 0, 255), radius=2):
+    for x, y in points.astype(np.int32):
+        cv2.circle(frame, (int(x), int(y)), radius, color, -1)
+
+
+if __name__ == '__main__':
+    import cv2
+
+    detector = RetinaFace(model_name=RetinaFaceWeights.MNET_050)
+    print(detector.get_info())
+    cap = cv2.VideoCapture(0)
+
+    if not cap.isOpened():
+        print('Failed to open webcam.')
+        exit()
+
+    print("Webcam started. Press 'q' to exit.")
+
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            print('Failed to read frame.')
+            break
+
+        # Get face detections as list of dictionaries
+        faces = detector.detect(frame)
+
+        # Process each detected face
+        for face in faces:
+            # Extract bbox and landmarks from dictionary
+            bbox = face['bbox']  # [x1, y1, x2, y2]
+            landmarks = face['landmarks']  # [[x1, y1], [x2, y2], ...]
+            confidence = face['confidence']
+
+            # Pass bbox and confidence separately
+            draw_bbox(frame, bbox, confidence)
+
+            # Convert landmarks to numpy array format if needed
+            if landmarks is not None and len(landmarks) > 0:
+                # Convert list of [x, y] pairs to numpy array
+                points = np.array(landmarks, dtype=np.float32)  # Shape: (5, 2)
+                draw_keypoints(frame, points)
+
+        # Display face count
+        cv2.putText(
+            frame,
+            f'Faces: {len(faces)}',
+            (10, 30),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.7,
+            (255, 255, 255),
+            2,
+        )
+
+        cv2.imshow('FaceDetection', frame)
+        if cv2.waitKey(1) & 0xFF == ord('q'):
+            break
+
+    cap.release()
+    cv2.destroyAllWindows()

uniface/detection/scrfd.py

@@ -0,0 +1,357 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from typing import Any, Dict, List, Literal, Tuple
+
+import cv2
+import numpy as np
+
+from uniface.common import distance2bbox, distance2kps, non_max_suppression, resize_image
+from uniface.constants import SCRFDWeights
+from uniface.log import Logger
+from uniface.model_store import verify_model_weights
+from uniface.onnx_utils import create_onnx_session
+
+from .base import BaseDetector
+
+__all__ = ['SCRFD']
+
+
+class SCRFD(BaseDetector):
+    """
+    Face detector based on the SCRFD architecture.
+
+    Title: "Sample and Computation Redistribution for Efficient Face Detection"
+    Paper: https://arxiv.org/abs/2105.04714
+    Code: https://github.com/insightface/insightface
+
+    Args:
+        model_name (SCRFDWeights): Predefined model enum (e.g., `SCRFD_10G_KPS`).
+            Specifies the SCRFD variant to load. Defaults to SCRFD_10G_KPS.
+        conf_thresh (float): Confidence threshold for filtering detections. Defaults to 0.5.
+        nms_thresh (float): Non-Maximum Suppression threshold. Defaults to 0.4.
+        input_size (Tuple[int, int]): Input image size (width, height).
+            Defaults to (640, 640).
+            Note: Non-default sizes may cause slower inference and CoreML compatibility issues.
+        **kwargs: Reserved for future advanced options.
+
+    Attributes:
+        model_name (SCRFDWeights): Selected model variant.
+        conf_thresh (float): Threshold used to filter low-confidence detections.
+        nms_thresh (float): Threshold used during NMS to suppress overlapping boxes.
+        input_size (Tuple[int, int]): Image size to which inputs are resized before inference.
+        _fmc (int): Number of feature map levels used in the model.
+        _feat_stride_fpn (List[int]): Feature map strides corresponding to each detection level.
+        _num_anchors (int): Number of anchors per feature location.
+        _center_cache (Dict): Cached anchor centers for efficient forward passes.
+        _model_path (str): Absolute path to the downloaded/verified model weights.
+
+    Raises:
+        ValueError: If the model weights are invalid or not found.
+        RuntimeError: If the ONNX model fails to load or initialize.
+    """
+
+    def __init__(
+        self,
+        *,
+        model_name: SCRFDWeights = SCRFDWeights.SCRFD_10G_KPS,
+        conf_thresh: float = 0.5,
+        nms_thresh: float = 0.4,
+        input_size: Tuple[int, int] = (640, 640),
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(
+            model_name=model_name,
+            conf_thresh=conf_thresh,
+            nms_thresh=nms_thresh,
+            input_size=input_size,
+            **kwargs,
+        )
+        self._supports_landmarks = True  # SCRFD supports landmarks
+
+        self.model_name = model_name
+        self.conf_thresh = conf_thresh
+        self.nms_thresh = nms_thresh
+        self.input_size = input_size
+
+        # ------- SCRFD model params ------
+        self._fmc = 3
+        self._feat_stride_fpn = [8, 16, 32]
+        self._num_anchors = 2
+        self._center_cache = {}
+        # ---------------------------------
+
+        Logger.info(
+            f'Initializing SCRFD with model={self.model_name}, conf_thresh={self.conf_thresh}, '
+            f'nms_thresh={self.nms_thresh}, input_size={self.input_size}'
+        )
+
+        # Get path to model weights
+        self._model_path = verify_model_weights(self.model_name)
+        Logger.info(f'Verified model weights located at: {self._model_path}')
+
+        # Initialize model
+        self._initialize_model(self._model_path)
+
+    def _initialize_model(self, model_path: str) -> None:
+        """
+        Initializes an ONNX model session from the given path.
+
+        Args:
+            model_path (str): The file path to the ONNX model.
+
+        Raises:
+            RuntimeError: If the model fails to load, logs an error and raises an exception.
+        """
+        try:
+            self.session = create_onnx_session(model_path)
+            self.input_names = self.session.get_inputs()[0].name
+            self.output_names = [x.name for x in self.session.get_outputs()]
+            Logger.info(f'Successfully initialized the model from {model_path}')
+        except Exception as e:
+            Logger.error(f"Failed to load model from '{model_path}': {e}", exc_info=True)
+            raise RuntimeError(f"Failed to initialize model session for '{model_path}'") from e
+
+    def preprocess(self, image: np.ndarray) -> Tuple[np.ndarray, Tuple[int, int]]:
+        """Preprocess image for inference.
+
+        Args:
+            image (np.ndarray): Input image
+
+        Returns:
+            Tuple[np.ndarray, Tuple[int, int]]: Preprocessed blob and input size
+        """
+        image = image.astype(np.float32)
+        image = (image - 127.5) / 127.5
+        image = image.transpose(2, 0, 1)  # HWC to CHW
+        image = np.expand_dims(image, axis=0)
+
+        return image
+
+    def inference(self, input_tensor: np.ndarray) -> List[np.ndarray]:
+        """Perform model inference on the preprocessed image tensor.
+
+        Args:
+            input_tensor (np.ndarray): Preprocessed input tensor.
+
+        Returns:
+            Tuple[np.ndarray, np.ndarray]: Raw model outputs.
+        """
+        return self.session.run(self.output_names, {self.input_names: input_tensor})
+
+    def postprocess(self, outputs: List[np.ndarray], image_size: Tuple[int, int]):
+        scores_list = []
+        bboxes_list = []
+        kpss_list = []
+
+        image_size = image_size
+
+        fmc = self._fmc
+        for idx, stride in enumerate(self._feat_stride_fpn):
+            scores = outputs[idx]
+            bbox_preds = outputs[fmc + idx] * stride
+            kps_preds = outputs[2 * fmc + idx] * stride
+
+            # Generate anchors
+            fm_height = image_size[0] // stride
+            fm_width = image_size[1] // stride
+            cache_key = (fm_height, fm_width, stride)
+
+            if cache_key in self._center_cache:
+                anchor_centers = self._center_cache[cache_key]
+            else:
+                y, x = np.mgrid[:fm_height, :fm_width]
+                anchor_centers = np.stack((x, y), axis=-1).astype(np.float32)
+                anchor_centers = (anchor_centers * stride).reshape(-1, 2)
+
+                if self._num_anchors > 1:
+                    anchor_centers = np.tile(anchor_centers[:, None, :], (1, self._num_anchors, 1)).reshape(-1, 2)
+
+                if len(self._center_cache) < 100:
+                    self._center_cache[cache_key] = anchor_centers
+
+            pos_indices = np.where(scores >= self.conf_thresh)[0]
+            if len(pos_indices) == 0:
+                continue
+
+            bboxes = distance2bbox(anchor_centers, bbox_preds)[pos_indices]
+            scores_selected = scores[pos_indices]
+            scores_list.append(scores_selected)
+            bboxes_list.append(bboxes)
+
+            landmarks = distance2kps(anchor_centers, kps_preds)
+            landmarks = landmarks.reshape((landmarks.shape[0], -1, 2))
+            kpss_list.append(landmarks[pos_indices])
+
+        return scores_list, bboxes_list, kpss_list
+
+    def detect(
+        self,
+        image: np.ndarray,
+        *,
+        max_num: int = 0,
+        metric: Literal['default', 'max'] = 'max',
+        center_weight: float = 2.0,
+    ) -> List[Dict[str, Any]]:
+        """
+        Perform face detection on an input image and return bounding boxes and facial landmarks.
+
+        Args:
+            image (np.ndarray): Input image as a NumPy array of shape (H, W, C).
+            max_num (int): Maximum number of detections to return. Use 0 to return all detections. Defaults to 0.
+            metric (Literal["default", "max"]): Metric for ranking detections when `max_num` is limited.
+                - "default": Prioritize detections closer to the image center.
+                - "max": Prioritize detections with larger bounding box areas.
+            center_weight (float): Weight for penalizing detections farther from the image center
+                when using the "default" metric. Defaults to 2.0.
+
+        Returns:
+            List[Dict[str, Any]]: List of face detection dictionaries, each containing:
+                - 'bbox' (np.ndarray): Bounding box coordinates with shape (4,) as [x1, y1, x2, y2]
+                - 'confidence' (float): Detection confidence score (0.0 to 1.0)
+                - 'landmarks' (np.ndarray): 5-point facial landmarks with shape (5, 2)
+
+        Example:
+            >>> faces = detector.detect(image)
+            >>> for face in faces:
+            ...     bbox = face['bbox']  # np.ndarray with shape (4,)
+            ...     confidence = face['confidence']  # float
+            ...     landmarks = face['landmarks']  # np.ndarray with shape (5, 2)
+            ...     # Can pass landmarks directly to recognition
+            ...     embedding = recognizer.get_normalized_embedding(image, landmarks)
+        """
+
+        original_height, original_width = image.shape[:2]
+
+        image, resize_factor = resize_image(image, target_shape=self.input_size)
+
+        image_tensor = self.preprocess(image)
+
+        # ONNXRuntime inference
+        outputs = self.inference(image_tensor)
+
+        scores_list, bboxes_list, kpss_list = self.postprocess(outputs, image_size=image.shape[:2])
+
+        # Handle case when no faces are detected
+        if not scores_list:
+            return []
+
+        scores = np.vstack(scores_list)
+        scores_ravel = scores.ravel()
+        order = scores_ravel.argsort()[::-1]
+
+        bboxes = np.vstack(bboxes_list) / resize_factor
+        landmarks = np.vstack(kpss_list) / resize_factor
+
+        pre_det = np.hstack((bboxes, scores)).astype(np.float32, copy=False)
+        pre_det = pre_det[order, :]
+
+        keep = non_max_suppression(pre_det, threshold=self.nms_thresh)
+
+        detections = pre_det[keep, :]
+        landmarks = landmarks[order, :, :]
+        landmarks = landmarks[keep, :, :].astype(np.float32)
+
+        if 0 < max_num < detections.shape[0]:
+            # Calculate area of detections
+            area = (detections[:, 2] - detections[:, 0]) * (detections[:, 3] - detections[:, 1])
+
+            # Calculate offsets from image center
+            center = (original_height // 2, original_width // 2)
+            offsets = np.vstack(
+                [
+                    (detections[:, 0] + detections[:, 2]) / 2 - center[1],
+                    (detections[:, 1] + detections[:, 3]) / 2 - center[0],
+                ]
+            )
+
+            # Calculate scores based on the chosen metric
+            offset_dist_squared = np.sum(np.power(offsets, 2.0), axis=0)
+            if metric == 'max':
+                values = area
+            else:
+                values = area - offset_dist_squared * center_weight
+
+            # Sort by scores and select top `max_num`
+            sorted_indices = np.argsort(values)[::-1][:max_num]
+            detections = detections[sorted_indices]
+            landmarks = landmarks[sorted_indices]
+
+        faces = []
+        for i in range(detections.shape[0]):
+            face_dict = {
+                'bbox': detections[i, :4],
+                'confidence': float(detections[i, 4]),
+                'landmarks': landmarks[i],
+            }
+            faces.append(face_dict)
+
+        return faces
+
+
+# TODO: below is only for testing, remove it later
+def draw_bbox(frame, bbox, score, color=(0, 255, 0), thickness=2):
+    x1, y1, x2, y2 = map(int, bbox)  # Unpack 4 bbox values
+    cv2.rectangle(frame, (x1, y1), (x2, y2), color, thickness)
+    cv2.putText(frame, f'{score:.2f}', (x1, y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, color, 1)
+
+
+def draw_keypoints(frame, points, color=(0, 0, 255), radius=2):
+    for x, y in points.astype(np.int32):
+        cv2.circle(frame, (int(x), int(y)), radius, color, -1)
+
+
+if __name__ == '__main__':
+    detector = SCRFD(model_name=SCRFDWeights.SCRFD_500M_KPS)
+    print(detector.get_info())
+    cap = cv2.VideoCapture(0)
+
+    if not cap.isOpened():
+        print('Failed to open webcam.')
+        exit()
+
+    print("Webcam started. Press 'q' to exit.")
+
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            print('Failed to read frame.')
+            break
+
+        # Get face detections as list of dictionaries
+        faces = detector.detect(frame)
+
+        # Process each detected face
+        for face in faces:
+            # Extract bbox and landmarks from dictionary
+            bbox = face['bbox']  # [x1, y1, x2, y2]
+            landmarks = face['landmarks']  # [[x1, y1], [x2, y2], ...]
+            confidence = face['confidence']
+
+            # Pass bbox and confidence separately
+            draw_bbox(frame, bbox, confidence)
+
+            # Convert landmarks to numpy array format if needed
+            if landmarks is not None and len(landmarks) > 0:
+                # Convert list of [x, y] pairs to numpy array
+                points = np.array(landmarks, dtype=np.float32)  # Shape: (5, 2)
+                draw_keypoints(frame, points)
+
+        # Display face count
+        cv2.putText(
+            frame,
+            f'Faces: {len(faces)}',
+            (10, 30),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.7,
+            (255, 255, 255),
+            2,
+        )
+
+        cv2.imshow('FaceDetection', frame)
+        if cv2.waitKey(1) & 0xFF == ord('q'):
+            break
+
+    cap.release()
+    cv2.destroyAllWindows()

uniface/detection/yolov5.py

@@ -0,0 +1,340 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from typing import Any, Dict, List, Literal, Tuple
+
+import cv2
+import numpy as np
+
+from uniface.common import non_max_suppression
+from uniface.constants import YOLOv5FaceWeights
+from uniface.log import Logger
+from uniface.model_store import verify_model_weights
+from uniface.onnx_utils import create_onnx_session
+
+from .base import BaseDetector
+
+__all__ = ['YOLOv5Face']
+
+
+class YOLOv5Face(BaseDetector):
+    """
+    Face detector based on the YOLOv5-Face architecture.
+
+    Title: "YOLO5Face: Why Reinventing a Face Detector"
+    Paper: https://arxiv.org/abs/2105.12931
+    Code: https://github.com/yakhyo/yolov5-face-onnx-inference (ONNX inference implementation)
+
+    Args:
+        model_name (YOLOv5FaceWeights): Predefined model enum (e.g., `YOLOV5S`).
+            Specifies the YOLOv5-Face variant to load. Defaults to YOLOV5S.
+        conf_thresh (float): Confidence threshold for filtering detections. Defaults to 0.6.
+        nms_thresh (float): Non-Maximum Suppression threshold. Defaults to 0.5.
+        input_size (int): Input image size. Defaults to 640.
+            Note: ONNX model is fixed at 640. Changing this will cause inference errors.
+        **kwargs: Advanced options:
+            max_det (int): Maximum number of detections to return. Defaults to 750.
+
+    Attributes:
+        model_name (YOLOv5FaceWeights): Selected model variant.
+        conf_thresh (float): Threshold used to filter low-confidence detections.
+        nms_thresh (float): Threshold used during NMS to suppress overlapping boxes.
+        input_size (int): Image size to which inputs are resized before inference.
+        max_det (int): Maximum number of detections to return.
+        _model_path (str): Absolute path to the downloaded/verified model weights.
+
+    Raises:
+        ValueError: If the model weights are invalid or not found.
+        RuntimeError: If the ONNX model fails to load or initialize.
+    """
+
+    def __init__(
+        self,
+        *,
+        model_name: YOLOv5FaceWeights = YOLOv5FaceWeights.YOLOV5S,
+        conf_thresh: float = 0.6,
+        nms_thresh: float = 0.5,
+        input_size: int = 640,
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(
+            model_name=model_name,
+            conf_thresh=conf_thresh,
+            nms_thresh=nms_thresh,
+            input_size=input_size,
+            **kwargs,
+        )
+        self._supports_landmarks = True  # YOLOv5-Face supports landmarks
+
+        # Validate input size
+        if input_size != 640:
+            raise ValueError(
+                f'YOLOv5Face only supports input_size=640 (got {input_size}). The ONNX model has a fixed input shape.'
+            )
+
+        self.model_name = model_name
+        self.conf_thresh = conf_thresh
+        self.nms_thresh = nms_thresh
+        self.input_size = input_size
+
+        # Advanced options from kwargs
+        self.max_det = kwargs.get('max_det', 750)
+
+        Logger.info(
+            f'Initializing YOLOv5Face with model={self.model_name}, conf_thresh={self.conf_thresh}, '
+            f'nms_thresh={self.nms_thresh}, input_size={self.input_size}'
+        )
+
+        # Get path to model weights
+        self._model_path = verify_model_weights(self.model_name)
+        Logger.info(f'Verified model weights located at: {self._model_path}')
+
+        # Initialize model
+        self._initialize_model(self._model_path)
+
+    def _initialize_model(self, model_path: str) -> None:
+        """
+        Initializes an ONNX model session from the given path.
+
+        Args:
+            model_path (str): The file path to the ONNX model.
+
+        Raises:
+            RuntimeError: If the model fails to load, logs an error and raises an exception.
+        """
+        try:
+            self.session = create_onnx_session(model_path)
+            self.input_names = self.session.get_inputs()[0].name
+            self.output_names = [x.name for x in self.session.get_outputs()]
+            Logger.info(f'Successfully initialized the model from {model_path}')
+        except Exception as e:
+            Logger.error(f"Failed to load model from '{model_path}': {e}", exc_info=True)
+            raise RuntimeError(f"Failed to initialize model session for '{model_path}'") from e
+
+    def preprocess(self, image: np.ndarray) -> Tuple[np.ndarray, float, Tuple[int, int]]:
+        """
+        Preprocess image for inference.
+
+        Args:
+            image (np.ndarray): Input image (BGR format)
+
+        Returns:
+            Tuple[np.ndarray, float, Tuple[int, int]]: Preprocessed image, scale ratio, and padding
+        """
+        # Get original image shape
+        img_h, img_w = image.shape[:2]
+
+        # Calculate scale ratio
+        scale = min(self.input_size / img_h, self.input_size / img_w)
+        new_h, new_w = int(img_h * scale), int(img_w * scale)
+
+        # Resize image
+        img_resized = cv2.resize(image, (new_w, new_h), interpolation=cv2.INTER_LINEAR)
+
+        # Create padded image
+        img_padded = np.full((self.input_size, self.input_size, 3), 114, dtype=np.uint8)
+
+        # Calculate padding
+        pad_h = (self.input_size - new_h) // 2
+        pad_w = (self.input_size - new_w) // 2
+
+        # Place resized image in center
+        img_padded[pad_h : pad_h + new_h, pad_w : pad_w + new_w] = img_resized
+
+        # Convert to RGB and normalize
+        img_rgb = cv2.cvtColor(img_padded, cv2.COLOR_BGR2RGB)
+        img_normalized = img_rgb.astype(np.float32) / 255.0
+
+        # Transpose to CHW format (HWC -> CHW) and add batch dimension
+        img_transposed = np.transpose(img_normalized, (2, 0, 1))
+        img_batch = np.expand_dims(img_transposed, axis=0)
+        img_batch = np.ascontiguousarray(img_batch)
+
+        return img_batch, scale, (pad_w, pad_h)
+
+    def inference(self, input_tensor: np.ndarray) -> List[np.ndarray]:
+        """Perform model inference on the preprocessed image tensor.
+
+        Args:
+            input_tensor (np.ndarray): Preprocessed input tensor.
+
+        Returns:
+            List[np.ndarray]: Raw model outputs.
+        """
+        return self.session.run(self.output_names, {self.input_names: input_tensor})
+
+    def postprocess(
+        self,
+        predictions: np.ndarray,
+        scale: float,
+        padding: Tuple[int, int],
+    ) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        Postprocess model predictions.
+
+        Args:
+            predictions (np.ndarray): Raw model output
+            scale (float): Scale ratio used in preprocessing
+            padding (Tuple[int, int]): Padding used in preprocessing
+
+        Returns:
+            Tuple[np.ndarray, np.ndarray]: Filtered detections and landmarks
+                - detections: [x1, y1, x2, y2, conf]
+                - landmarks: [5, 2] for each detection
+        """
+        # predictions shape: (1, 25200, 16)
+        # 16 = [x, y, w, h, obj_conf, cls_conf, 10 landmarks (5 points * 2 coords)]
+
+        predictions = predictions[0]  # Remove batch dimension
+
+        # Filter by confidence
+        mask = predictions[:, 4] >= self.conf_thresh
+        predictions = predictions[mask]
+
+        if len(predictions) == 0:
+            return np.array([]), np.array([])
+
+        # Convert from xywh to xyxy
+        boxes = self._xywh2xyxy(predictions[:, :4])
+
+        # Get confidence scores
+        scores = predictions[:, 4]
+
+        # Get landmarks (5 points, 10 coordinates)
+        landmarks = predictions[:, 5:15].copy()
+
+        # Apply NMS
+        detections_for_nms = np.hstack((boxes, scores[:, None])).astype(np.float32, copy=False)
+        keep = non_max_suppression(detections_for_nms, self.nms_thresh)
+
+        if len(keep) == 0:
+            return np.array([]), np.array([])
+
+        # Filter detections and limit to max_det
+        keep = keep[: self.max_det]
+        boxes = boxes[keep]
+        scores = scores[keep]
+        landmarks = landmarks[keep]
+
+        # Scale back to original image coordinates
+        pad_w, pad_h = padding
+        boxes[:, [0, 2]] = (boxes[:, [0, 2]] - pad_w) / scale
+        boxes[:, [1, 3]] = (boxes[:, [1, 3]] - pad_h) / scale
+
+        # Scale landmarks
+        for i in range(5):
+            landmarks[:, i * 2] = (landmarks[:, i * 2] - pad_w) / scale
+            landmarks[:, i * 2 + 1] = (landmarks[:, i * 2 + 1] - pad_h) / scale
+
+        # Reshape landmarks to (N, 5, 2)
+        landmarks = landmarks.reshape(-1, 5, 2)
+
+        # Combine results
+        detections = np.concatenate([boxes, scores[:, None]], axis=1)
+
+        return detections, landmarks
+
+    def _xywh2xyxy(self, x: np.ndarray) -> np.ndarray:
+        """
+        Convert bounding box format from xywh to xyxy.
+
+        Args:
+            x (np.ndarray): Boxes in [x, y, w, h] format
+
+        Returns:
+            np.ndarray: Boxes in [x1, y1, x2, y2] format
+        """
+        y = np.copy(x)
+        y[..., 0] = x[..., 0] - x[..., 2] / 2  # x1
+        y[..., 1] = x[..., 1] - x[..., 3] / 2  # y1
+        y[..., 2] = x[..., 0] + x[..., 2] / 2  # x2
+        y[..., 3] = x[..., 1] + x[..., 3] / 2  # y2
+        return y
+
+    def detect(
+        self,
+        image: np.ndarray,
+        *,
+        max_num: int = 0,
+        metric: Literal['default', 'max'] = 'max',
+        center_weight: float = 2.0,
+    ) -> List[Dict[str, Any]]:
+        """
+        Perform face detection on an input image and return bounding boxes and facial landmarks.
+
+        Args:
+            image (np.ndarray): Input image as a NumPy array of shape (H, W, C).
+            max_num (int): Maximum number of detections to return. Use 0 to return all detections. Defaults to 0.
+            metric (Literal["default", "max"]): Metric for ranking detections when `max_num` is limited.
+                - "default": Prioritize detections closer to the image center.
+                - "max": Prioritize detections with larger bounding box areas.
+            center_weight (float): Weight for penalizing detections farther from the image center
+                when using the "default" metric. Defaults to 2.0.
+
+        Returns:
+            List[Dict[str, Any]]: List of face detection dictionaries, each containing:
+                - 'bbox' (np.ndarray): Bounding box coordinates with shape (4,) as [x1, y1, x2, y2]
+                - 'confidence' (float): Detection confidence score (0.0 to 1.0)
+                - 'landmarks' (np.ndarray): 5-point facial landmarks with shape (5, 2)
+
+        Example:
+            >>> faces = detector.detect(image)
+            >>> for face in faces:
+            ...     bbox = face['bbox']  # np.ndarray with shape (4,)
+            ...     confidence = face['confidence']  # float
+            ...     landmarks = face['landmarks']  # np.ndarray with shape (5, 2)
+            ...     # Can pass landmarks directly to recognition
+            ...     embedding = recognizer.get_normalized_embedding(image, landmarks)
+        """
+
+        original_height, original_width = image.shape[:2]
+
+        # Preprocess
+        image_tensor, scale, padding = self.preprocess(image)
+
+        # ONNXRuntime inference
+        outputs = self.inference(image_tensor)
+
+        # Postprocess
+        detections, landmarks = self.postprocess(outputs[0], scale, padding)
+
+        # Handle case when no faces are detected
+        if len(detections) == 0:
+            return []
+
+        if 0 < max_num < detections.shape[0]:
+            # Calculate area of detections
+            area = (detections[:, 2] - detections[:, 0]) * (detections[:, 3] - detections[:, 1])
+
+            # Calculate offsets from image center
+            center = (original_height // 2, original_width // 2)
+            offsets = np.vstack(
+                [
+                    (detections[:, 0] + detections[:, 2]) / 2 - center[1],
+                    (detections[:, 1] + detections[:, 3]) / 2 - center[0],
+                ]
+            )
+
+            # Calculate scores based on the chosen metric
+            offset_dist_squared = np.sum(np.power(offsets, 2.0), axis=0)
+            if metric == 'max':
+                values = area
+            else:
+                values = area - offset_dist_squared * center_weight
+
+            # Sort by scores and select top `max_num`
+            sorted_indices = np.argsort(values)[::-1][:max_num]
+            detections = detections[sorted_indices]
+            landmarks = landmarks[sorted_indices]
+
+        faces = []
+        for i in range(detections.shape[0]):
+            face_dict = {
+                'bbox': detections[i, :4],
+                'confidence': float(detections[i, 4]),
+                'landmarks': landmarks[i],
+            }
+            faces.append(face_dict)
+
+        return faces

uniface/face.py

@@ -0,0 +1,66 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from dataclasses import asdict, dataclass
+from typing import Optional
+
+import numpy as np
+
+from uniface.face_utils import compute_similarity
+
+__all__ = ['Face']
+
+
+@dataclass
+class Face:
+    """
+    Detected face with analysis results.
+    """
+
+    # Required attributes
+    bbox: np.ndarray
+    confidence: float
+    landmarks: np.ndarray
+
+    # Optional attributes
+    embedding: Optional[np.ndarray] = None
+    age: Optional[int] = None
+    gender: Optional[int] = None  # 0 or 1
+
+    def compute_similarity(self, other: 'Face') -> float:
+        """Compute cosine similarity with another face."""
+        if self.embedding is None or other.embedding is None:
+            raise ValueError('Both faces must have embeddings for similarity computation')
+        return float(compute_similarity(self.embedding, other.embedding))
+
+    def to_dict(self) -> dict:
+        """Convert to dictionary."""
+        return asdict(self)
+
+    @property
+    def sex(self) -> str:
+        """Get gender as a string label (Female or Male)."""
+        if self.gender is None:
+            return None
+        return 'Female' if self.gender == 0 else 'Male'
+
+    @property
+    def bbox_xyxy(self) -> np.ndarray:
+        """Get bounding box coordinates in (x1, y1, x2, y2) format."""
+        return self.bbox.copy()
+
+    @property
+    def bbox_xywh(self) -> np.ndarray:
+        """Get bounding box coordinates in (x1, y1, w, h) format."""
+        return np.array([self.bbox[0], self.bbox[1], self.bbox[2] - self.bbox[0], self.bbox[3] - self.bbox[1]])
+
+    def __repr__(self) -> str:
+        parts = [f'Face(confidence={self.confidence:.3f}']
+        if self.age is not None:
+            parts.append(f'age={self.age}')
+        if self.gender is not None:
+            parts.append(f'sex={self.sex}')
+        if self.embedding is not None:
+            parts.append(f'embedding_dim={self.embedding.shape[0]}')
+        return ', '.join(parts) + ')'

uniface/face_utils.py

@@ -0,0 +1,194 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from typing import Tuple, Union
+
+import cv2
+import numpy as np
+from skimage.transform import SimilarityTransform
+
+__all__ = [
+    'face_alignment',
+    'compute_similarity',
+    'bbox_center_alignment',
+    'transform_points_2d',
+]
+
+
+# Reference alignment for facial landmarks (ArcFace)
+reference_alignment: np.ndarray = np.array(
+    [
+        [38.2946, 51.6963],
+        [73.5318, 51.5014],
+        [56.0252, 71.7366],
+        [41.5493, 92.3655],
+        [70.7299, 92.2041],
+    ],
+    dtype=np.float32,
+)
+
+
+def estimate_norm(landmark: np.ndarray, image_size: Union[int, Tuple[int, int]] = 112) -> Tuple[np.ndarray, np.ndarray]:
+    """
+    Estimate the normalization transformation matrix for facial landmarks.
+
+    Args:
+        landmark (np.ndarray): Array of shape (5, 2) representing the coordinates of the facial landmarks.
+        image_size (Union[int, Tuple[int, int]], optional): The size of the output image.
+            Can be an integer (for square images) or a tuple (width, height). Default is 112.
+
+    Returns:
+        np.ndarray: The 2x3 transformation matrix for aligning the landmarks.
+        np.ndarray: The 2x3 inverse transformation matrix for aligning the landmarks.
+
+    Raises:
+        AssertionError: If the input landmark array does not have the shape (5, 2)
+                        or if image_size is not a multiple of 112 or 128.
+    """
+    assert landmark.shape == (5, 2), 'Landmark array must have shape (5, 2).'
+
+    # Handle both int and tuple inputs
+    if isinstance(image_size, tuple):
+        size = image_size[0]  # Use width for ratio calculation
+    else:
+        size = image_size
+
+    assert size % 112 == 0 or size % 128 == 0, 'Image size must be a multiple of 112 or 128.'
+
+    if size % 112 == 0:
+        ratio = float(size) / 112.0
+        diff_x = 0.0
+    else:
+        ratio = float(size) / 128.0
+        diff_x = 8.0 * ratio
+
+    # Adjust reference alignment based on ratio and diff_x
+    alignment = reference_alignment * ratio
+    alignment[:, 0] += diff_x
+
+    # Compute the transformation matrix
+    transform = SimilarityTransform()
+    transform.estimate(landmark, alignment)
+
+    matrix = transform.params[0:2, :]
+    inverse_matrix = np.linalg.inv(transform.params)[0:2, :]
+
+    return matrix, inverse_matrix
+
+
+def face_alignment(
+    image: np.ndarray,
+    landmark: np.ndarray,
+    image_size: Union[int, Tuple[int, int]] = 112,
+) -> Tuple[np.ndarray, np.ndarray]:
+    """
+    Align the face in the input image based on the given facial landmarks.
+
+    Args:
+        image (np.ndarray): Input image as a NumPy array.
+        landmark (np.ndarray): Array of shape (5, 2) representing the coordinates of the facial landmarks.
+        image_size (Union[int, Tuple[int, int]], optional): The size of the aligned output image.
+            Can be an integer (for square images) or a tuple (width, height). Default is 112.
+
+    Returns:
+        np.ndarray: The aligned face as a NumPy array.
+        np.ndarray: The 2x3 transformation matrix used for alignment.
+    """
+    # Get the transformation matrix
+    M, M_inv = estimate_norm(landmark, image_size)
+
+    # Handle both int and tuple for warpAffine output size
+    if isinstance(image_size, int):
+        output_size = (image_size, image_size)
+    else:
+        output_size = image_size
+
+    # Warp the input image to align the face
+    warped = cv2.warpAffine(image, M, output_size, borderValue=0.0)
+
+    return warped, M_inv
+
+
+def compute_similarity(feat1: np.ndarray, feat2: np.ndarray, normalized: bool = False) -> np.float32:
+    """Computing Similarity between two faces.
+
+    Args:
+        feat1 (np.ndarray): First embedding.
+        feat2 (np.ndarray): Second embedding.
+        normalized (bool): Set True if the embeddings are already L2 normalized.
+
+    Returns:
+        np.float32: Cosine similarity.
+    """
+    feat1 = feat1.ravel()
+    feat2 = feat2.ravel()
+    if normalized:
+        return np.dot(feat1, feat2)
+    else:
+        return np.dot(feat1, feat2) / (np.linalg.norm(feat1) * np.linalg.norm(feat2) + 1e-5)
+
+
+def bbox_center_alignment(image, center, output_size, scale, rotation):
+    """
+    Apply center-based alignment, scaling, and rotation to an image.
+
+    Args:
+        image (np.ndarray): Input image.
+        center (Tuple[float, float]): Center point (e.g., face center from bbox).
+        output_size (int): Desired output image size (square).
+        scale (float): Scaling factor to zoom in/out.
+        rotation (float): Rotation angle in degrees (clockwise).
+
+    Returns:
+        cropped (np.ndarray): Aligned and cropped image.
+        M (np.ndarray): 2x3 affine transform matrix used.
+    """
+
+    # Convert rotation from degrees to radians
+    rot = float(rotation) * np.pi / 180.0
+
+    # Scale the image
+    t1 = SimilarityTransform(scale=scale)
+
+    # Translate the center point to the origin (after scaling)
+    cx = center[0] * scale
+    cy = center[1] * scale
+    t2 = SimilarityTransform(translation=(-1 * cx, -1 * cy))
+
+    # Apply rotation around origin (center of face)
+    t3 = SimilarityTransform(rotation=rot)
+
+    # Translate origin to center of output image
+    t4 = SimilarityTransform(translation=(output_size / 2, output_size / 2))
+
+    # Combine all transformations in order: scale → center shift → rotate → recentralize
+    t = t1 + t2 + t3 + t4
+
+    # Extract 2x3 affine matrix
+    M = t.params[0:2]
+
+    # Warp the image using OpenCV
+    cropped = cv2.warpAffine(image, M, (output_size, output_size), borderValue=0.0)
+
+    return cropped, M
+
+
+def transform_points_2d(points: np.ndarray, transform: np.ndarray) -> np.ndarray:
+    """
+    Apply a 2D affine transformation to an array of 2D points.
+
+    Args:
+        points (np.ndarray): An (N, 2) array of 2D points.
+        transform (np.ndarray): A (2, 3) affine transformation matrix.
+
+    Returns:
+        np.ndarray: Transformed (N, 2) array of points.
+    """
+    transformed = np.zeros_like(points, dtype=np.float32)
+    for i in range(points.shape[0]):
+        point = np.array([points[i, 0], points[i, 1], 1.0], dtype=np.float32)
+        result = np.dot(transform, point)
+        transformed[i] = result[:2]
+
+    return transformed

uniface/gaze/__init__.py

@@ -0,0 +1,58 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from .base import BaseGazeEstimator
+from .models import MobileGaze
+
+
+def create_gaze_estimator(method: str = 'mobilegaze', **kwargs) -> BaseGazeEstimator:
+    """
+    Factory function to create gaze estimators.
+
+    This function initializes and returns a gaze estimator instance based on the
+    specified method. It acts as a high-level interface to the underlying
+    model classes.
+
+    Args:
+        method (str): The gaze estimation method to use.
+            Options: 'mobilegaze' (default).
+        **kwargs: Model-specific parameters passed to the estimator's constructor.
+            For example, `model_name` can be used to select a specific
+            backbone from `GazeWeights` enum (RESNET18, RESNET34, RESNET50,
+            MOBILENET_V2, MOBILEONE_S0).
+
+    Returns:
+        BaseGazeEstimator: An initialized gaze estimator instance ready for use.
+
+    Raises:
+        ValueError: If the specified `method` is not supported.
+
+    Examples:
+        >>> # Create the default MobileGaze estimator (ResNet18 backbone)
+        >>> estimator = create_gaze_estimator()
+
+        >>> # Create with MobileNetV2 backbone
+        >>> from uniface.constants import GazeWeights
+        >>> estimator = create_gaze_estimator(
+        ...     'mobilegaze',
+        ...     model_name=GazeWeights.MOBILENET_V2
+        ... )
+
+        >>> # Use the estimator
+        >>> pitch, yaw = estimator.estimate(face_crop)
+    """
+    method = method.lower()
+
+    if method in ('mobilegaze', 'mobile_gaze', 'gaze'):
+        return MobileGaze(**kwargs)
+    else:
+        available = ['mobilegaze']
+        raise ValueError(f"Unsupported gaze estimation method: '{method}'. Available: {available}")
+
+
+__all__ = [
+    'create_gaze_estimator',
+    'MobileGaze',
+    'BaseGazeEstimator',
+]

uniface/gaze/base.py

@@ -0,0 +1,108 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from abc import ABC, abstractmethod
+from typing import Tuple
+
+import numpy as np
+
+
+class BaseGazeEstimator(ABC):
+    """
+    Abstract base class for all gaze estimation models.
+
+    This class defines the common interface that all gaze estimators must implement,
+    ensuring consistency across different gaze estimation methods. Gaze estimation
+    predicts the direction a person is looking based on their face image.
+
+    The gaze direction is represented as pitch and yaw angles in radians:
+    - Pitch: Vertical angle (positive = looking up, negative = looking down)
+    - Yaw: Horizontal angle (positive = looking right, negative = looking left)
+    """
+
+    @abstractmethod
+    def _initialize_model(self) -> None:
+        """
+        Initialize the underlying model for inference.
+
+        This method should handle loading model weights, creating the
+        inference session (e.g., ONNX Runtime), and any necessary
+        setup procedures to prepare the model for prediction.
+
+        Raises:
+            RuntimeError: If the model fails to load or initialize.
+        """
+        raise NotImplementedError('Subclasses must implement the _initialize_model method.')
+
+    @abstractmethod
+    def preprocess(self, face_image: np.ndarray) -> np.ndarray:
+        """
+        Preprocess the input face image for model inference.
+
+        This method should take a raw face crop and convert it into the format
+        expected by the model's inference engine (e.g., normalized tensor).
+
+        Args:
+            face_image (np.ndarray): A cropped face image in BGR format with
+                                     shape (H, W, C).
+
+        Returns:
+            np.ndarray: The preprocessed image tensor ready for inference,
+                        typically with shape (1, C, H, W).
+        """
+        raise NotImplementedError('Subclasses must implement the preprocess method.')
+
+    @abstractmethod
+    def postprocess(self, outputs: Tuple[np.ndarray, np.ndarray]) -> Tuple[float, float]:
+        """
+        Postprocess raw model outputs into gaze angles.
+
+        This method takes the raw output from the model's inference and
+        converts it into pitch and yaw angles in radians.
+
+        Args:
+            outputs: Raw outputs from the model inference. The format depends
+                     on the specific model architecture.
+
+        Returns:
+            Tuple[float, float]: A tuple of (pitch, yaw) angles in radians.
+        """
+        raise NotImplementedError('Subclasses must implement the postprocess method.')
+
+    @abstractmethod
+    def estimate(self, face_image: np.ndarray) -> Tuple[float, float]:
+        """
+        Perform end-to-end gaze estimation on a face image.
+
+        This method orchestrates the full pipeline: preprocessing the input,
+        running inference, and postprocessing to return the gaze direction.
+
+        Args:
+            face_image (np.ndarray): A cropped face image in BGR format.
+                                     The face should be roughly centered and
+                                     well-framed within the image.
+
+        Returns:
+            Tuple[float, float]: A tuple of (pitch, yaw) angles in radians:
+                - pitch: Vertical gaze angle (positive = up, negative = down)
+                - yaw: Horizontal gaze angle (positive = right, negative = left)
+
+        Example:
+            >>> estimator = create_gaze_estimator()
+            >>> pitch, yaw = estimator.estimate(face_crop)
+            >>> print(f"Looking: pitch={np.degrees(pitch):.1f}°, yaw={np.degrees(yaw):.1f}°")
+        """
+        raise NotImplementedError('Subclasses must implement the estimate method.')
+
+    def __call__(self, face_image: np.ndarray) -> Tuple[float, float]:
+        """
+        Provides a convenient, callable shortcut for the `estimate` method.
+
+        Args:
+            face_image (np.ndarray): A cropped face image in BGR format.
+
+        Returns:
+            Tuple[float, float]: A tuple of (pitch, yaw) angles in radians.
+        """
+        return self.estimate(face_image)

uniface/gaze/models.py

@@ -0,0 +1,187 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from typing import Tuple
+
+import cv2
+import numpy as np
+
+from uniface.constants import GazeWeights
+from uniface.log import Logger
+from uniface.model_store import verify_model_weights
+from uniface.onnx_utils import create_onnx_session
+
+from .base import BaseGazeEstimator
+
+__all__ = ['MobileGaze']
+
+
+class MobileGaze(BaseGazeEstimator):
+    """
+    MobileGaze: Real-Time Gaze Estimation with ONNX Runtime.
+
+    MobileGaze is a gaze estimation model that predicts gaze direction from a single
+    face image. It supports multiple backbone architectures including ResNet 18/34/50,
+    MobileNetV2, and MobileOne S0. The model uses a classification approach with binned
+    angles, which are then decoded to continuous pitch and yaw values.
+
+    The model outputs gaze direction as pitch (vertical) and yaw (horizontal) angles
+    in radians.
+
+    Reference:
+        https://github.com/yakhyo/gaze-estimation
+
+    Args:
+        model_name (GazeWeights): The enum specifying the gaze model backbone to load.
+            Options: RESNET18, RESNET34, RESNET50, MOBILENET_V2, MOBILEONE_S0.
+            Defaults to `GazeWeights.RESNET18`.
+        input_size (Tuple[int, int]): The resolution (width, height) for the model's
+            input. Defaults to (448, 448).
+
+    Attributes:
+        input_size (Tuple[int, int]): Model input dimensions.
+        input_mean (list): Per-channel mean values for normalization (ImageNet).
+        input_std (list): Per-channel std values for normalization (ImageNet).
+
+    Example:
+        >>> from uniface.gaze import MobileGaze
+        >>> from uniface import RetinaFace
+        >>>
+        >>> detector = RetinaFace()
+        >>> gaze_estimator = MobileGaze()
+        >>>
+        >>> # Detect faces and estimate gaze for each
+        >>> faces = detector.detect(image)
+        >>> for face in faces:
+        ...     bbox = face['bbox']
+        ...     x1, y1, x2, y2 = map(int, bbox[:4])
+        ...     face_crop = image[y1:y2, x1:x2]
+        ...     pitch, yaw = gaze_estimator.estimate(face_crop)
+        ...     print(f"Gaze: pitch={np.degrees(pitch):.1f}°, yaw={np.degrees(yaw):.1f}°")
+    """
+
+    def __init__(
+        self,
+        model_name: GazeWeights = GazeWeights.RESNET34,
+        input_size: Tuple[int, int] = (448, 448),
+    ) -> None:
+        Logger.info(f'Initializing MobileGaze with model={model_name}, input_size={input_size}')
+
+        self.input_size = input_size
+        self.input_mean = [0.485, 0.456, 0.406]
+        self.input_std = [0.229, 0.224, 0.225]
+
+        # Model specific parameters for bin-based classification (Gaze360 config)
+        self._bins = 90
+        self._binwidth = 4
+        self._angle_offset = 180
+        self._idx_tensor = np.arange(self._bins, dtype=np.float32)
+
+        self.model_path = verify_model_weights(model_name)
+        self._initialize_model()
+
+    def _initialize_model(self) -> None:
+        """
+        Initialize the ONNX model from the stored model path.
+
+        Raises:
+            RuntimeError: If the model fails to load or initialize.
+        """
+        try:
+            self.session = create_onnx_session(self.model_path)
+
+            # Get input configuration
+            input_cfg = self.session.get_inputs()[0]
+            input_shape = input_cfg.shape
+            self.input_name = input_cfg.name
+            self.input_size = tuple(input_shape[2:4][::-1])  # Update from model
+
+            # Get output configuration
+            outputs = self.session.get_outputs()
+            self.output_names = [output.name for output in outputs]
+
+            if len(self.output_names) != 2:
+                raise ValueError(f'Expected 2 output nodes (pitch, yaw), got {len(self.output_names)}')
+
+            Logger.info(f'MobileGaze initialized with input size {self.input_size}')
+
+        except Exception as e:
+            Logger.error(f"Failed to load gaze model from '{self.model_path}'", exc_info=True)
+            raise RuntimeError(f'Failed to initialize gaze model: {e}') from e
+
+    def preprocess(self, face_image: np.ndarray) -> np.ndarray:
+        """
+        Preprocess a face crop for gaze estimation.
+
+        Args:
+            face_image (np.ndarray): A cropped face image in BGR format.
+
+        Returns:
+            np.ndarray: Preprocessed image tensor with shape (1, 3, H, W).
+        """
+        # Convert BGR to RGB
+        image = cv2.cvtColor(face_image, cv2.COLOR_BGR2RGB)
+
+        # Resize to model input size
+        image = cv2.resize(image, self.input_size)
+
+        # Normalize to [0, 1] and apply normalization
+        image = image.astype(np.float32) / 255.0
+        mean = np.array(self.input_mean, dtype=np.float32)
+        std = np.array(self.input_std, dtype=np.float32)
+        image = (image - mean) / std
+
+        # HWC -> CHW -> NCHW
+        image = np.transpose(image, (2, 0, 1))
+        image = np.expand_dims(image, axis=0).astype(np.float32)
+
+        return image
+
+    def _softmax(self, x: np.ndarray) -> np.ndarray:
+        """Apply softmax along axis 1."""
+        e_x = np.exp(x - np.max(x, axis=1, keepdims=True))
+        return e_x / e_x.sum(axis=1, keepdims=True)
+
+    def postprocess(self, outputs: Tuple[np.ndarray, np.ndarray]) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        Postprocess raw model outputs into gaze angles.
+
+        This method takes the raw output from the model's inference and
+        converts it into pitch and yaw angles in radians.
+
+        Args:
+            outputs: Raw outputs from the model inference. The format depends
+                     on the specific model architecture.
+
+        Returns:
+            Tuple[np.ndarray, np.ndarray]: A tuple of (pitch, yaw) angles in radians.
+        """
+        pitch_logits, yaw_logits = outputs
+
+        # Convert logits to probabilities
+        pitch_probs = self._softmax(pitch_logits)
+        yaw_probs = self._softmax(yaw_logits)
+
+        # Compute expected bin index (soft-argmax)
+        pitch_deg = np.sum(pitch_probs * self._idx_tensor, axis=1) * self._binwidth - self._angle_offset
+        yaw_deg = np.sum(yaw_probs * self._idx_tensor, axis=1) * self._binwidth - self._angle_offset
+
+        # Convert degrees to radians
+        pitch = np.radians(pitch_deg[0])
+        yaw = np.radians(yaw_deg[0])
+
+        return pitch, yaw
+
+    def estimate(self, face_image: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        Perform end-to-end gaze estimation on a face image.
+
+        This method orchestrates the full pipeline: preprocessing the input,
+        running inference, and postprocessing to return the gaze direction.
+        """
+        input_tensor = self.preprocess(face_image)
+        outputs = self.session.run(self.output_names, {self.input_name: input_tensor})
+        pitch, yaw = self.postprocess((outputs[0], outputs[1]))
+
+        return pitch, yaw

uniface/landmark/__init__.py

@@ -0,0 +1,28 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from .base import BaseLandmarker
+from .models import Landmark106
+
+
+def create_landmarker(method: str = '2d106det', **kwargs) -> BaseLandmarker:
+    """
+    Factory function to create facial landmark predictors.
+
+    Args:
+        method (str): Landmark prediction method. Options: '106'.
+        **kwargs: Model-specific parameters.
+
+    Returns:
+        Initialized landmarker instance.
+    """
+    method = method.lower()
+    if method == '2d106det':
+        return Landmark106(**kwargs)
+    else:
+        available = ['2d106det']
+        raise ValueError(f"Unsupported method: '{method}'. Available: {available}")
+
+
+__all__ = ['create_landmarker', 'Landmark106', 'BaseLandmarker']

uniface/landmark/base.py

@@ -0,0 +1,32 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from abc import ABC, abstractmethod
+
+import numpy as np
+
+
+class BaseLandmarker(ABC):
+    """
+    Abstract Base Class for all facial landmark models.
+    """
+
+    @abstractmethod
+    def get_landmarks(self, image: np.ndarray, bbox: np.ndarray) -> np.ndarray:
+        """
+        Predicts facial landmarks for a given face bounding box.
+
+        This method defines the standard interface for all landmark predictors.
+        It takes a full image and a bounding box for a single face and returns
+        the predicted keypoints for that face.
+
+        Args:
+            image (np.ndarray): The full source image in BGR format.
+            bbox (np.ndarray): A bounding box of a face [x1, y1, x2, y2].
+
+        Returns:
+            np.ndarray: An array of predicted landmark points with shape (N, 2),
+                        where N is the number of landmarks.
+        """
+        raise NotImplementedError

uniface/landmark/models.py

@@ -0,0 +1,212 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from typing import Tuple
+
+import cv2
+import numpy as np
+
+from uniface.constants import LandmarkWeights
+from uniface.face_utils import bbox_center_alignment, transform_points_2d
+from uniface.log import Logger
+from uniface.model_store import verify_model_weights
+from uniface.onnx_utils import create_onnx_session
+
+from .base import BaseLandmarker
+
+__all__ = ['Landmark106']
+
+
+class Landmark106(BaseLandmarker):
+    """Facial landmark model for predicting 106 facial keypoints.
+
+    This class implements the BaseLandmarker and provides an end-to-end
+    pipeline for 106-point facial landmark detection. It handles model
+    loading, preprocessing of a face crop based on a bounding box,
+    inference, and post-processing to map landmarks back to the
+    original image coordinates.
+
+    Args:
+        model_name (LandmarkWeights): The enum specifying the landmark model to load.
+            Defaults to `LandmarkWeights.DEFAULT`.
+        input_size (Tuple[int, int]): The resolution (width, height) for the model's
+            input. Defaults to (192, 192).
+
+    Example:
+        >>> # Assume 'image' is a loaded image and 'bbox' is a face bounding box
+        >>> # bbox = [x1, y1, x2, y2]
+        >>>
+        >>> landmarker = Landmark106()
+        >>> landmarks = landmarker.get_landmarks(image, bbox)
+        >>> print(landmarks.shape)
+        (106, 2)
+    """
+
+    def __init__(
+        self,
+        model_name: LandmarkWeights = LandmarkWeights.DEFAULT,
+        input_size: Tuple[int, int] = (192, 192),
+    ) -> None:
+        Logger.info(f'Initializing Facial Landmark with model={model_name}, input_size={input_size}')
+        self.input_size = input_size
+        self.input_std = 1.0
+        self.input_mean = 0.0
+        self.model_path = verify_model_weights(model_name)
+        self._initialize_model()
+
+    def _initialize_model(self):
+        """
+        Initialize the ONNX model from the stored model path.
+
+        Raises:
+            RuntimeError: If the model fails to load or initialize.
+        """
+        try:
+            self.session = create_onnx_session(self.model_path)
+
+            # Get input configuration
+            input_metadata = self.session.get_inputs()[0]
+            input_shape = input_metadata.shape
+            self.input_size = tuple(input_shape[2:4][::-1])  # Update input size from model
+
+            # Get input/output names
+            self.input_names = [input.name for input in self.session.get_inputs()]
+            self.output_names = [output.name for output in self.session.get_outputs()]
+
+            # Determine landmark dimensions from output shape
+            output_shape = self.session.get_outputs()[0].shape
+            self.lmk_dim = 2  # x,y coordinates
+            self.lmk_num = output_shape[1] // self.lmk_dim  # Number of landmarks
+
+            Logger.info(f'Model initialized with {self.lmk_num} landmarks')
+
+        except Exception as e:
+            Logger.error(f"Failed to load landmark model from '{self.model_path}'", exc_info=True)
+            raise RuntimeError(f'Failed to initialize landmark model: {e}') from e
+
+    def preprocess(self, image: np.ndarray, bbox: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
+        """Prepares a face crop for inference.
+
+        This method takes a face bounding box, performs a center alignment to
+        warp the face into the model's required input size, and then creates
+        a normalized blob ready for the ONNX session.
+
+        Args:
+            image (np.ndarray): The full source image in BGR format.
+            bbox (np.ndarray): The bounding box of the face [x1, y1, x2, y2].
+
+        Returns:
+            Tuple[np.ndarray, np.ndarray]: A tuple containing:
+                - The preprocessed image blob ready for inference.
+                - The affine transformation matrix used for alignment.
+        """
+        width, height = bbox[2] - bbox[0], bbox[3] - bbox[1]
+        center = ((bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2)
+        scale = self.input_size[0] / (max(width, height) * 1.5)
+
+        aligned_face, transform_matrix = bbox_center_alignment(image, center, self.input_size[0], scale, 0.0)
+
+        face_blob = cv2.dnn.blobFromImage(
+            aligned_face,
+            1.0 / self.input_std,
+            self.input_size,
+            (self.input_mean, self.input_mean, self.input_mean),
+            swapRB=True,
+        )
+        return face_blob, transform_matrix
+
+    def postprocess(self, predictions: np.ndarray, transform_matrix: np.ndarray) -> np.ndarray:
+        """Converts raw model predictions back to original image coordinates.
+
+        This method reshapes the model's flat output array into landmark points,
+        denormalizes them to the model's input space, and then applies an
+        inverse affine transformation to map them back to the original image space.
+
+        Args:
+            predictions (np.ndarray): Raw landmark coordinates from the model output.
+            transform_matrix (np.ndarray): The affine transformation matrix from preprocessing.
+
+        Returns:
+            np.ndarray: An array of landmark points in the original image's coordinates.
+        """
+        landmarks = predictions.reshape((-1, 2))
+        landmarks[:, 0:2] += 1
+        landmarks[:, 0:2] *= self.input_size[0] // 2
+
+        inverse_matrix = cv2.invertAffineTransform(transform_matrix)
+        landmarks = transform_points_2d(landmarks, inverse_matrix)
+        return landmarks
+
+    def get_landmarks(self, image: np.ndarray, bbox: np.ndarray) -> np.ndarray:
+        """Predicts facial landmarks for the given image and face bounding box.
+
+        This is the main public method that orchestrates the full pipeline of
+        preprocessing, inference, and post-processing.
+
+        Args:
+            image (np.ndarray): The full source image in BGR format.
+            bbox (np.ndarray): A bounding box of a face [x1, y1, x2, y2].
+
+        Returns:
+            np.ndarray: An array of predicted landmark points with shape (106, 2).
+        """
+        face_blob, transform_matrix = self.preprocess(image, bbox)
+        raw_predictions = self.session.run(self.output_names, {self.input_names[0]: face_blob})[0][0]
+        landmarks = self.postprocess(raw_predictions, transform_matrix)
+        return landmarks
+
+
+# Testing code
+if __name__ == '__main__':
+    from uniface.detection import RetinaFace
+    from uniface.landmark import Landmark106
+
+    face_detector = RetinaFace()
+    landmarker = Landmark106()
+
+    cap = cv2.VideoCapture(0)
+    if not cap.isOpened():
+        print('Webcam not available.')
+        exit()
+
+    print("Press 'q' to quit.")
+
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            print('Frame capture failed.')
+            break
+
+        # 2. The detect method returns a list of dictionaries
+        faces = face_detector.detect(frame)
+
+        if not faces:
+            cv2.imshow('Facial Landmark Detection', frame)
+            if cv2.waitKey(1) & 0xFF == ord('q'):
+                break
+            continue
+
+        # 3. Loop through the list of face dictionaries
+        for face in faces:
+            # Extract the bounding box
+            bbox = face['bbox']
+
+            # 4. Get landmarks for the current face using its bounding box
+            landmarks = landmarker.get_landmarks(frame, bbox)
+
+            # --- Drawing Logic ---
+            # Draw the landmarks
+            for x, y in landmarks.astype(int):
+                cv2.circle(frame, (x, y), 2, (0, 255, 0), -1)
+
+            # Draw the bounding box
+            x1, y1, x2, y2 = map(int, bbox)
+            cv2.rectangle(frame, (x1, y1), (x2, y2), (255, 0, 0), 2)
+
+        cv2.imshow('Facial Landmark Detection', frame)
+        if cv2.waitKey(1) & 0xFF == ord('q'):
+            break
+
+    cap.release()
+    cv2.destroyAllWindows()

uniface/log.py

@@ -1,7 +1,25 @@
 import logging
 
-logging.basicConfig(
-    level=logging.INFO,
-    format="%(asctime)s - %(levelname)s - %(message)s"
-)
-Logger = logging.getLogger("retinaface")
+# Create logger for uniface
+Logger = logging.getLogger('uniface')
+Logger.setLevel(logging.WARNING)  # Only show warnings/errors by default
+Logger.addHandler(logging.NullHandler())
+
+
+def enable_logging(level=logging.INFO):
+    """
+    Enable verbose logging for uniface.
+
+    Args:
+        level: Logging level (logging.DEBUG, logging.INFO, etc.)
+
+    Example:
+        >>> from uniface import enable_logging
+        >>> enable_logging()  # Show INFO logs
+    """
+    Logger.handlers.clear()
+    handler = logging.StreamHandler()
+    handler.setFormatter(logging.Formatter('%(asctime)s - %(levelname)s - %(message)s', datefmt='%Y-%m-%d %H:%M:%S'))
+    Logger.addHandler(handler)
+    Logger.setLevel(level)
+    Logger.propagate = False

uniface/model_store.py

@@ -1,102 +1,115 @@
-# Copyright 2024 Yakhyokhuja Valikhujaev
-# Author: Yakhyokhuja Valikhujaev
-# GitHub: https://github.com/yakhyo
-
-import os
-import hashlib
-import requests
-
-from uniface.log import Logger
-import uniface.constants as const
-
-
-def verify_model_weights(model_name: str, root: str = '~/.uniface/models') -> str:
-    """
-    Ensures model weights are available by downloading if missing and verifying integrity with a SHA-256 hash.
-
-    Checks if the specified model weights file exists in `root`. If missing, downloads from a predefined URL.
-    The file is then verified using its SHA-256 hash. If verification fails, the corrupted file is deleted,
-    and an error is raised.
-
-    Args:
-        model_name (str): Name of the model weights to verify or download.
-        root (str, optional): Directory to store the model weights. Defaults to '~/.uniface/models'.
-
-    Returns:
-        str: Path to the verified model weights file.
-
-    Raises:
-        ValueError: If the model is not found or if verification fails.
-        ConnectionError: If downloading the file fails.
-
-    Examples:
-        >>> # Download and verify 'retinaface_mnet025' weights
-        >>> verify_model_weights('retinaface_mnet025')
-        '/home/user/.uniface/models/retinaface_mnet025.onnx'
-
-        >>> # Use a custom directory
-        >>> verify_model_weights('retinaface_r34', root='/custom/dir')
-        '/custom/dir/retinaface_r34.onnx'
-    """
-
-    root = os.path.expanduser(root)
-    os.makedirs(root, exist_ok=True)
-    model_path = os.path.join(root, f'{model_name}.onnx')
-
-    if not os.path.exists(model_path):
-        url = const.MODEL_URLS.get(model_name)
-        if not url:
-            Logger.error(f"No URL found for model '{model_name}'")
-            raise ValueError(f"No URL found for model '{model_name}'")
-
-        Logger.info(f"Downloading '{model_name}' from {url}")
-        download_file(url, model_path)
-        Logger.info(f"Successfully '{model_name}' downloaded to {model_path}")
-
-    expected_hash = const.MODEL_SHA256.get(model_name)
-    if expected_hash and not verify_file_hash(model_path, expected_hash):
-        os.remove(model_path)  # Remove corrupted file
-        Logger.warning("Corrupted weight detected. Removing...")
-        raise ValueError(f"Hash mismatch for '{model_name}'. The file may be corrupted; please try downloading again.")
-
-    return model_path
-
-
-def download_file(url: str, dest_path: str) -> None:
-    """Download a file from a URL in chunks and save it to the destination path."""
-    try:
-        response = requests.get(url, stream=True)
-        response.raise_for_status()
-        with open(dest_path, "wb") as file:
-            for chunk in response.iter_content(chunk_size=const.CHUNK_SIZE):
-                if chunk:
-                    file.write(chunk)
-    except requests.RequestException as e:
-        raise ConnectionError(f"Failed to download file from {url}. Error: {e}")
-
-
-def verify_file_hash(file_path: str, expected_hash: str) -> bool:
-    """Compute the SHA-256 hash of the file and compare it with the expected hash."""
-    file_hash = hashlib.sha256()
-    with open(file_path, "rb") as f:
-        for chunk in iter(lambda: f.read(const.CHUNK_SIZE), b""):
-            file_hash.update(chunk)
-    actual_hash = file_hash.hexdigest()
-    if actual_hash != expected_hash:
-        Logger.warning(f"Expected hash: {expected_hash}, but got: {actual_hash}")
-    return actual_hash == expected_hash
-
-
-if __name__ == "__main__":
-    model_names = [
-        'retinaface_mnet025',
-        'retinaface_mnet050',
-        'retinaface_mnet_v1',
-        'retinaface_mnet_v2',
-        'retinaface_r18',
-        'retinaface_r34'
-    ]
-
-    # Download each model in the list
-    for model_name in model_names:
-        model_path = verify_model_weights(model_name)
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+import hashlib
+import os
+
+import requests
+from tqdm import tqdm
+
+import uniface.constants as const
+from uniface.log import Logger
+
+__all__ = ['verify_model_weights']
+
+
+def verify_model_weights(model_name: str, root: str = '~/.uniface/models') -> str:
+    """
+    Ensure model weights are present, downloading and verifying them using SHA-256 if necessary.
+
+    Given a model identifier from an Enum class (e.g., `RetinaFaceWeights.MNET_V2`), this function checks if
+    the corresponding `.onnx` weight file exists locally. If not, it downloads the file from a predefined URL.
+    After download, the file’s integrity is verified using a SHA-256 hash. If verification fails, the file is deleted
+    and an error is raised.
+
+    Args:
+        model_name (Enum): Model weight identifier (e.g., `RetinaFaceWeights.MNET_V2`, `ArcFaceWeights.RESNET`, etc.).
+        root (str, optional): Directory to store or locate the model weights. Defaults to '~/.uniface/models'.
+
+    Returns:
+        str: Absolute path to the verified model weights file.
+
+    Raises:
+        ValueError: If the model is unknown or SHA-256 verification fails.
+        ConnectionError: If downloading the file fails.
+
+    Examples:
+        >>> from uniface.models import RetinaFaceWeights, verify_model_weights
+        >>> verify_model_weights(RetinaFaceWeights.MNET_V2)
+        '/home/user/.uniface/models/retinaface_mnet_v2.onnx'
+
+        >>> verify_model_weights(RetinaFaceWeights.RESNET34, root='/custom/dir')
+        '/custom/dir/retinaface_r34.onnx'
+    """
+
+    root = os.path.expanduser(root)
+    os.makedirs(root, exist_ok=True)
+
+    # Keep model_name as enum for dictionary lookup
+    url = const.MODEL_URLS.get(model_name)
+    if not url:
+        Logger.error(f"No URL found for model '{model_name}'")
+        raise ValueError(f"No URL found for model '{model_name}'")
+
+    file_ext = os.path.splitext(url)[1]
+    model_path = os.path.normpath(os.path.join(root, f'{model_name.value}{file_ext}'))
+
+    if not os.path.exists(model_path):
+        Logger.info(f"Downloading model '{model_name}' from {url}")
+        try:
+            download_file(url, model_path)
+            Logger.info(f"Successfully downloaded '{model_name}' to {model_path}")
+        except Exception as e:
+            Logger.error(f"Failed to download model '{model_name}': {e}")
+            raise ConnectionError(f"Download failed for '{model_name}'") from e
+
+    expected_hash = const.MODEL_SHA256.get(model_name)
+    if expected_hash and not verify_file_hash(model_path, expected_hash):
+        os.remove(model_path)  # Remove corrupted file
+        Logger.warning('Corrupted weight detected. Removing...')
+        raise ValueError(f"Hash mismatch for '{model_name}'. The file may be corrupted; please try downloading again.")
+
+    return model_path
+
+
+def download_file(url: str, dest_path: str) -> None:
+    """Download a file from a URL in chunks and save it to the destination path."""
+    try:
+        response = requests.get(url, stream=True)
+        response.raise_for_status()
+        with (
+            open(dest_path, 'wb') as file,
+            tqdm(
+                desc=f'Downloading {dest_path}',
+                unit='B',
+                unit_scale=True,
+                unit_divisor=1024,
+            ) as progress,
+        ):
+            for chunk in response.iter_content(chunk_size=const.CHUNK_SIZE):
+                if chunk:
+                    file.write(chunk)
+                    progress.update(len(chunk))
+    except requests.RequestException as e:
+        raise ConnectionError(f'Failed to download file from {url}. Error: {e}') from e
+
+
+def verify_file_hash(file_path: str, expected_hash: str) -> bool:
+    """Compute the SHA-256 hash of the file and compare it with the expected hash."""
+    file_hash = hashlib.sha256()
+    with open(file_path, 'rb') as f:
+        for chunk in iter(lambda: f.read(const.CHUNK_SIZE), b''):
+            file_hash.update(chunk)
+    actual_hash = file_hash.hexdigest()
+    if actual_hash != expected_hash:
+        Logger.warning(f'Expected hash: {expected_hash}, but got: {actual_hash}')
+    return actual_hash == expected_hash
+
+
+if __name__ == '__main__':
+    model_names = [model.value for model in const.RetinaFaceWeights]
+
+    # Download each model in the list
+    for model_name in model_names:
+        model_path = verify_model_weights(model_name)

uniface/onnx_utils.py

@@ -0,0 +1,99 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from typing import List
+
+import onnxruntime as ort
+
+from uniface.log import Logger
+
+
+def get_available_providers() -> List[str]:
+    """
+    Get list of available ONNX Runtime execution providers for the current platform.
+
+    Automatically detects and prioritizes hardware acceleration:
+    - CoreML on Apple Silicon (M1/M2/M3/M4)
+    - CUDA on NVIDIA GPUs
+    - CPU as fallback (always available)
+
+    Returns:
+        List[str]: Ordered list of execution providers to use
+
+    Examples:
+        >>> providers = get_available_providers()
+        >>> # On M4 Mac: ['CoreMLExecutionProvider', 'CPUExecutionProvider']
+        >>> # On Linux with CUDA: ['CUDAExecutionProvider', 'CPUExecutionProvider']
+        >>> # On CPU-only: ['CPUExecutionProvider']
+    """
+    available = ort.get_available_providers()
+    providers = []
+
+    # Priority order: CoreML > CUDA > CPU
+    if 'CoreMLExecutionProvider' in available:
+        providers.append('CoreMLExecutionProvider')
+        Logger.info('CoreML acceleration enabled (Apple Silicon)')
+
+    if 'CUDAExecutionProvider' in available:
+        providers.append('CUDAExecutionProvider')
+        Logger.info('CUDA acceleration enabled (NVIDIA GPU)')
+
+    # CPU is always available as fallback
+    providers.append('CPUExecutionProvider')
+
+    if len(providers) == 1:
+        Logger.info('Using CPU execution (no hardware acceleration detected)')
+
+    return providers
+
+
+def create_onnx_session(model_path: str, providers: List[str] = None) -> ort.InferenceSession:
+    """
+    Create an ONNX Runtime inference session with optimal provider selection.
+
+    Args:
+        model_path (str): Path to the ONNX model file
+        providers (List[str], optional): List of providers to use.
+            If None, automatically detects best available providers.
+
+    Returns:
+        ort.InferenceSession: Configured ONNX Runtime session
+
+    Raises:
+        RuntimeError: If session creation fails
+
+    Examples:
+        >>> session = create_onnx_session("model.onnx")
+        >>> # Automatically uses best available providers
+
+        >>> session = create_onnx_session("model.onnx", providers=["CPUExecutionProvider"])
+        >>> # Force CPU-only execution
+    """
+    if providers is None:
+        providers = get_available_providers()
+
+    # Suppress ONNX Runtime warnings (e.g., CoreML partition warnings)
+    # Log levels: 0=VERBOSE, 1=INFO, 2=WARNING, 3=ERROR, 4=FATAL
+    sess_options = ort.SessionOptions()
+    sess_options.log_severity_level = 3  # Only show ERROR and FATAL
+
+    try:
+        session = ort.InferenceSession(model_path, sess_options=sess_options, providers=providers)
+        active_provider = session.get_providers()[0]
+        Logger.debug(f'Session created with provider: {active_provider}')
+
+        # Show user-friendly message about which provider is being used
+        provider_names = {
+            'CoreMLExecutionProvider': 'CoreML (Apple Silicon)',
+            'CUDAExecutionProvider': 'CUDA (NVIDIA GPU)',
+            'CPUExecutionProvider': 'CPU',
+        }
+        provider_display = provider_names.get(active_provider, active_provider)
+        Logger.debug(f'Model loaded with provider: {active_provider}')
+        print(f'✓ Model loaded ({provider_display})')
+
+        return session
+    except Exception as e:
+        Logger.error(f'Failed to create ONNX session: {e}', exc_info=True)
+        raise RuntimeError(f'Failed to initialize ONNX Runtime session: {e}') from e

uniface/parsing/__init__.py

@@ -0,0 +1,61 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from typing import Union
+
+from uniface.constants import ParsingWeights
+
+from .base import BaseFaceParser
+from .bisenet import BiSeNet
+
+__all__ = ['BaseFaceParser', 'BiSeNet', 'create_face_parser']
+
+
+def create_face_parser(
+    model_name: Union[str, ParsingWeights] = ParsingWeights.RESNET18,
+) -> BaseFaceParser:
+    """
+    Factory function to create a face parsing model instance.
+
+    This function provides a convenient way to instantiate face parsing models
+    without directly importing the specific model classes. It supports both
+    string-based and enum-based model selection.
+
+    Args:
+        model_name (Union[str, ParsingWeights]): The face parsing model to create.
+            Can be either a string or a ParsingWeights enum value.
+            Available options:
+            - 'parsing_resnet18' or ParsingWeights.RESNET18 (default)
+            - 'parsing_resnet34' or ParsingWeights.RESNET34
+
+    Returns:
+        BaseFaceParser: An instance of the requested face parsing model.
+
+    Raises:
+        ValueError: If the model_name is not recognized.
+
+    Examples:
+        >>> # Using enum
+        >>> from uniface.parsing import create_face_parser
+        >>> from uniface.constants import ParsingWeights
+        >>> parser = create_face_parser(ParsingWeights.RESNET18)
+        >>>
+        >>> # Using string
+        >>> parser = create_face_parser('parsing_resnet18')
+        >>>
+        >>> # Parse a face image
+        >>> mask = parser.parse(face_crop)
+    """
+    # Convert string to enum if necessary
+    if isinstance(model_name, str):
+        try:
+            model_name = ParsingWeights(model_name)
+        except ValueError as e:
+            valid_models = [e.value for e in ParsingWeights]
+            raise ValueError(
+                f"Unknown face parsing model: '{model_name}'. Valid options are: {', '.join(valid_models)}"
+            ) from e
+
+    # All parsing models use the same BiSeNet class
+    return BiSeNet(model_name=model_name)

uniface/parsing/base.py

@@ -0,0 +1,106 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from abc import ABC, abstractmethod
+from typing import Tuple
+
+import numpy as np
+
+
+class BaseFaceParser(ABC):
+    """
+    Abstract base class for all face parsing models.
+
+    This class defines the common interface that all face parsing models must implement,
+    ensuring consistency across different parsing methods. Face parsing segments a face
+    image into semantic regions such as skin, eyes, nose, mouth, hair, etc.
+
+    The output is a segmentation mask where each pixel is assigned a class label
+    representing a facial component.
+    """
+
+    @abstractmethod
+    def _initialize_model(self) -> None:
+        """
+        Initialize the underlying model for inference.
+
+        This method should handle loading model weights, creating the
+        inference session (e.g., ONNX Runtime), and any necessary
+        setup procedures to prepare the model for prediction.
+
+        Raises:
+            RuntimeError: If the model fails to load or initialize.
+        """
+        raise NotImplementedError('Subclasses must implement the _initialize_model method.')
+
+    @abstractmethod
+    def preprocess(self, face_image: np.ndarray) -> np.ndarray:
+        """
+        Preprocess the input face image for model inference.
+
+        This method should take a raw face crop and convert it into the format
+        expected by the model's inference engine (e.g., normalized tensor).
+
+        Args:
+            face_image (np.ndarray): A face image in BGR format with
+                                     shape (H, W, C).
+
+        Returns:
+            np.ndarray: The preprocessed image tensor ready for inference,
+                        typically with shape (1, C, H, W).
+        """
+        raise NotImplementedError('Subclasses must implement the preprocess method.')
+
+    @abstractmethod
+    def postprocess(self, outputs: np.ndarray, original_size: Tuple[int, int]) -> np.ndarray:
+        """
+        Postprocess raw model outputs into a segmentation mask.
+
+        This method takes the raw output from the model's inference and
+        converts it into a segmentation mask at the original image size.
+
+        Args:
+            outputs (np.ndarray): Raw outputs from the model inference.
+            original_size (Tuple[int, int]): Original image size (width, height).
+
+        Returns:
+            np.ndarray: Segmentation mask with the same size as the original image.
+        """
+        raise NotImplementedError('Subclasses must implement the postprocess method.')
+
+    @abstractmethod
+    def parse(self, face_image: np.ndarray) -> np.ndarray:
+        """
+        Perform end-to-end face parsing on a face image.
+
+        This method orchestrates the full pipeline: preprocessing the input,
+        running inference, and postprocessing to return the segmentation mask.
+
+        Args:
+            face_image (np.ndarray): A face image in BGR format.
+                                     The face should be roughly centered and
+                                     well-framed within the image.
+
+        Returns:
+            np.ndarray: Segmentation mask with the same size as input image,
+                       where each pixel value represents a facial component class.
+
+        Example:
+            >>> parser = create_face_parser()
+            >>> mask = parser.parse(face_crop)
+            >>> print(f"Mask shape: {mask.shape}, unique classes: {np.unique(mask)}")
+        """
+        raise NotImplementedError('Subclasses must implement the parse method.')
+
+    def __call__(self, face_image: np.ndarray) -> np.ndarray:
+        """
+        Provides a convenient, callable shortcut for the `parse` method.
+
+        Args:
+            face_image (np.ndarray): A face image in BGR format.
+
+        Returns:
+            np.ndarray: Segmentation mask with the same size as input image.
+        """
+        return self.parse(face_image)

uniface/parsing/bisenet.py

@@ -0,0 +1,166 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from typing import Tuple
+
+import cv2
+import numpy as np
+
+from uniface.constants import ParsingWeights
+from uniface.log import Logger
+from uniface.model_store import verify_model_weights
+from uniface.onnx_utils import create_onnx_session
+
+from .base import BaseFaceParser
+
+__all__ = ['BiSeNet']
+
+
+class BiSeNet(BaseFaceParser):
+    """
+    BiSeNet: Bilateral Segmentation Network for Face Parsing with ONNX Runtime.
+
+    BiSeNet is a semantic segmentation model that segments a face image into
+    different facial components such as skin, eyes, nose, mouth, hair, etc. The model
+    uses a BiSeNet architecture with ResNet backbone and outputs a segmentation mask
+    where each pixel is assigned a class label.
+
+    The model supports 19 facial component classes including:
+    - Background, skin, eyebrows, eyes, nose, mouth, lips, ears, hair, etc.
+
+    Reference:
+        https://github.com/yakhyo/face-parsing
+
+    Args:
+        model_name (ParsingWeights): The enum specifying the parsing model to load.
+            Options: RESNET18, RESNET34.
+            Defaults to `ParsingWeights.RESNET18`.
+        input_size (Tuple[int, int]): The resolution (width, height) for the model's
+            input. Defaults to (512, 512).
+
+    Attributes:
+        input_size (Tuple[int, int]): Model input dimensions.
+        input_mean (np.ndarray): Per-channel mean values for normalization (ImageNet).
+        input_std (np.ndarray): Per-channel std values for normalization (ImageNet).
+
+    Example:
+        >>> from uniface.parsing import BiSeNet
+        >>> from uniface import RetinaFace
+        >>>
+        >>> detector = RetinaFace()
+        >>> parser = BiSeNet()
+        >>>
+        >>> # Detect faces and parse each face
+        >>> faces = detector.detect(image)
+        >>> for face in faces:
+        ...     bbox = face['bbox']
+        ...     x1, y1, x2, y2 = map(int, bbox[:4])
+        ...     face_crop = image[y1:y2, x1:x2]
+        ...     mask = parser.parse(face_crop)
+        ...     print(f"Mask shape: {mask.shape}, unique classes: {np.unique(mask)}")
+    """
+
+    def __init__(
+        self,
+        model_name: ParsingWeights = ParsingWeights.RESNET18,
+        input_size: Tuple[int, int] = (512, 512),
+    ) -> None:
+        Logger.info(f'Initializing BiSeNet with model={model_name}, input_size={input_size}')
+
+        self.input_size = input_size
+        self.input_mean = np.array([0.485, 0.456, 0.406], dtype=np.float32)
+        self.input_std = np.array([0.229, 0.224, 0.225], dtype=np.float32)
+
+        self.model_path = verify_model_weights(model_name)
+        self._initialize_model()
+
+    def _initialize_model(self) -> None:
+        """
+        Initialize the ONNX model from the stored model path.
+
+        Raises:
+            RuntimeError: If the model fails to load or initialize.
+        """
+        try:
+            self.session = create_onnx_session(self.model_path)
+
+            # Get input configuration
+            input_cfg = self.session.get_inputs()[0]
+            input_shape = input_cfg.shape
+            self.input_name = input_cfg.name
+            self.input_size = tuple(input_shape[2:4][::-1])  # Update from model
+
+            # Get output configuration
+            outputs = self.session.get_outputs()
+            self.output_names = [output.name for output in outputs]
+
+            Logger.info(f'BiSeNet initialized with input size {self.input_size}')
+
+        except Exception as e:
+            Logger.error(f"Failed to load parsing model from '{self.model_path}'", exc_info=True)
+            raise RuntimeError(f'Failed to initialize parsing model: {e}') from e
+
+    def preprocess(self, face_image: np.ndarray) -> np.ndarray:
+        """
+        Preprocess a face image for parsing.
+
+        Args:
+            face_image (np.ndarray): A face image in BGR format.
+
+        Returns:
+            np.ndarray: Preprocessed image tensor with shape (1, 3, H, W).
+        """
+        # Convert BGR to RGB
+        image = cv2.cvtColor(face_image, cv2.COLOR_BGR2RGB)
+
+        # Resize to model input size
+        image = cv2.resize(image, self.input_size, interpolation=cv2.INTER_LINEAR)
+
+        # Normalize to [0, 1] and apply normalization
+        image = image.astype(np.float32) / 255.0
+        image = (image - self.input_mean) / self.input_std
+
+        # HWC -> CHW -> NCHW
+        image = np.transpose(image, (2, 0, 1))
+        image = np.expand_dims(image, axis=0).astype(np.float32)
+
+        return image
+
+    def postprocess(self, outputs: np.ndarray, original_size: Tuple[int, int]) -> np.ndarray:
+        """
+        Postprocess model output to segmentation mask.
+
+        Args:
+            outputs (np.ndarray): Raw model output.
+            original_size (Tuple[int, int]): Original image size (width, height).
+
+        Returns:
+            np.ndarray: Segmentation mask resized to original dimensions.
+        """
+        # Get the class with highest probability for each pixel
+        predicted_mask = outputs.squeeze(0).argmax(0).astype(np.uint8)
+
+        # Resize back to original size
+        restored_mask = cv2.resize(predicted_mask, original_size, interpolation=cv2.INTER_NEAREST)
+
+        return restored_mask
+
+    def parse(self, face_image: np.ndarray) -> np.ndarray:
+        """
+        Perform end-to-end face parsing on a face image.
+
+        This method orchestrates the full pipeline: preprocessing the input,
+        running inference, and postprocessing to return the segmentation mask.
+
+        Args:
+            face_image (np.ndarray): A face image in BGR format.
+
+        Returns:
+            np.ndarray: Segmentation mask with the same size as input image.
+        """
+        original_size = (face_image.shape[1], face_image.shape[0])  # (width, height)
+        input_tensor = self.preprocess(face_image)
+        outputs = self.session.run(self.output_names, {self.input_name: input_tensor})
+
+        return self.postprocess(outputs[0], original_size)

uniface/recognition/__init__.py

@@ -0,0 +1,64 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+
+from .base import BaseRecognizer
+from .models import ArcFace, MobileFace, SphereFace
+
+
+def create_recognizer(method: str = 'arcface', **kwargs) -> BaseRecognizer:
+    """
+    Factory function to create face recognizers.
+
+    This function initializes and returns a face recognizer instance based on the
+    specified method. It acts as a high-level interface to the underlying
+    model classes like ArcFace, MobileFace, etc.
+
+    Args:
+        method (str): The recognition method to use.
+            Options: 'arcface' (default), 'mobileface', 'sphereface'.
+        **kwargs: Model-specific parameters passed to the recognizer's constructor.
+            For example, `model_name` can be used to select a specific
+            pre-trained weight from the available enums (e.g., `ArcFaceWeights.MNET`).
+
+    Returns:
+        BaseRecognizer: An initialized recognizer instance ready for use.
+
+    Raises:
+        ValueError: If the specified `method` is not supported.
+
+    Examples:
+        >>> # Create the default ArcFace recognizer
+        >>> recognizer = create_recognizer()
+
+        >>> # Create a specific MobileFace recognizer
+        >>> from uniface.constants import MobileFaceWeights
+        >>> recognizer = create_recognizer(
+        ...     'mobileface',
+        ...     model_name=MobileFaceWeights.MNET_V2
+        ... )
+
+        >>> # Create a SphereFace recognizer
+        >>> recognizer = create_recognizer('sphereface')
+    """
+    method = method.lower()
+
+    if method == 'arcface':
+        return ArcFace(**kwargs)
+    elif method == 'mobileface':
+        return MobileFace(**kwargs)
+    elif method == 'sphereface':
+        return SphereFace(**kwargs)
+    else:
+        available = ['arcface', 'mobileface', 'sphereface']
+        raise ValueError(f"Unsupported method: '{method}'. Available: {available}")
+
+
+__all__ = [
+    'create_recognizer',
+    'ArcFace',
+    'MobileFace',
+    'SphereFace',
+    'BaseRecognizer',
+]

uniface/recognition/base.py

@@ -0,0 +1,156 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from typing import List, Tuple, Union
+
+import cv2
+import numpy as np
+
+from uniface.face_utils import face_alignment
+from uniface.log import Logger
+from uniface.onnx_utils import create_onnx_session
+
+
+@dataclass
+class PreprocessConfig:
+    """
+    Configuration for preprocessing images before feeding them into the model.
+    """
+
+    input_mean: Union[float, List[float]] = 127.5
+    input_std: Union[float, List[float]] = 127.5
+    input_size: Tuple[int, int] = (112, 112)
+
+
+class BaseRecognizer(ABC):
+    """
+    Abstract Base Class for all face recognition models.
+    It provides the core functionality for preprocessing, inference, and embedding extraction.
+    """
+
+    @abstractmethod
+    def __init__(self, model_path: str, preprocessing: PreprocessConfig) -> None:
+        """
+        Initializes the model. Subclasses must call this.
+
+        Args:
+            model_path (str): The direct path to the verified ONNX model.
+            preprocessing (PreprocessConfig): The configuration for preprocessing.
+        """
+        self.input_mean = preprocessing.input_mean
+        self.input_std = preprocessing.input_std
+        self.input_size = preprocessing.input_size
+
+        self.model_path = model_path
+        self._initialize_model()
+
+    def _initialize_model(self) -> None:
+        """
+        Loads the ONNX model and prepares it for inference.
+
+        Raises:
+            RuntimeError: If the model fails to load or initialize.
+        """
+        try:
+            # Initialize model session with available providers
+            self.session = create_onnx_session(self.model_path)
+
+            # Extract input configuration
+            input_cfg = self.session.get_inputs()[0]
+            self.input_name = input_cfg.name
+
+            # Verify input dimensions match our configuration
+            input_shape = input_cfg.shape
+            model_input_size = tuple(input_shape[2:4][::-1])  # (width, height)
+            if model_input_size != self.input_size:
+                Logger.warning(f'Model input size {model_input_size} differs from configured size {self.input_size}')
+
+            # Extract output configuration
+            self.output_names = [output.name for output in self.session.get_outputs()]
+            self.output_shape = self.session.get_outputs()[0].shape
+
+            assert len(self.output_names) == 1, 'Expected only one output node.'
+            Logger.info(f'Successfully initialized face encoder from {self.model_path}')
+
+        except Exception as e:
+            Logger.error(
+                f"Failed to load face encoder model from '{self.model_path}'",
+                exc_info=True,
+            )
+            raise RuntimeError(f"Failed to initialize model session for '{self.model_path}'") from e
+
+    def preprocess(self, face_img: np.ndarray) -> np.ndarray:
+        """
+        Preprocess the image: resize, normalize, and convert it to a blob.
+
+        Args:
+            face_img: Input image in BGR format.
+
+        Returns:
+            Preprocessed image as a NumPy array ready for inference.
+        """
+        resized_img = cv2.resize(face_img, self.input_size)
+
+        if isinstance(self.input_std, (list, tuple)):
+            # Per-channel normalization
+            rgb_img = cv2.cvtColor(resized_img, cv2.COLOR_BGR2RGB).astype(np.float32)
+            normalized_img = (rgb_img - np.array(self.input_mean, dtype=np.float32)) / np.array(
+                self.input_std, dtype=np.float32
+            )
+
+            # Change to NCHW (batch, channels, height, width)
+            blob = np.transpose(normalized_img, (2, 0, 1))  # CHW
+            blob = np.expand_dims(blob, axis=0)  # NCHW
+        else:
+            # Single-value normalization
+            blob = cv2.dnn.blobFromImage(
+                resized_img,
+                scalefactor=1.0 / self.input_std,
+                size=self.input_size,
+                mean=(self.input_mean, self.input_mean, self.input_mean),
+                swapRB=True,  # Convert BGR to RGB
+            )
+
+        return blob
+
+    def get_embedding(self, image: np.ndarray, landmarks: np.ndarray = None) -> np.ndarray:
+        """
+        Extracts face embedding from an image.
+
+        Args:
+            image: Input face image (BGR format). If already aligned (112x112), landmarks can be None.
+            landmarks: Facial landmarks (5 points for alignment). Optional if image is already aligned.
+
+        Returns:
+            Face embedding vector (typically 512-dimensional).
+        """
+        # If landmarks are provided, align the face first
+        if landmarks is not None:
+            aligned_face, _ = face_alignment(image, landmarks, image_size=self.input_size)
+        else:
+            # Assume image is already aligned
+            aligned_face = image
+
+        # Generate embedding from aligned face
+        face_blob = self.preprocess(aligned_face)
+        embedding = self.session.run(self.output_names, {self.input_name: face_blob})[0]
+
+        return embedding
+
+    def get_normalized_embedding(self, image: np.ndarray, landmarks: np.ndarray) -> np.ndarray:
+        """
+        Extracts a l2 normalized face embedding vector from an image.
+
+        Args:
+            image: Input face image (BGR format).
+            landmarks: Facial landmarks (5 points for alignment).
+
+        Returns:
+            Normalized face embedding vector (typically 512-dimensional).
+        """
+        embedding = self.get_embedding(image, landmarks)
+        norm = np.linalg.norm(embedding)
+        return embedding / norm if norm > 0 else embedding

uniface/recognition/models.py

@@ -0,0 +1,103 @@
+# Copyright 2025 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from typing import Optional
+
+from uniface.constants import ArcFaceWeights, MobileFaceWeights, SphereFaceWeights
+from uniface.model_store import verify_model_weights
+
+from .base import BaseRecognizer, PreprocessConfig
+
+__all__ = ['ArcFace', 'MobileFace', 'SphereFace']
+
+
+class ArcFace(BaseRecognizer):
+    """ArcFace model for robust face recognition.
+
+    This class provides a concrete implementation of the BaseRecognizer,
+    pre-configured for ArcFace models. It handles the loading of specific
+    ArcFace weights and sets up the appropriate default preprocessing.
+
+    Args:
+        model_name (ArcFaceWeights): The specific ArcFace model variant to use.
+            Defaults to `ArcFaceWeights.MNET`.
+        preprocessing (Optional[PreprocessConfig]): An optional custom preprocessing
+            configuration. If None, a default config for ArcFace is used.
+
+    Example:
+        >>> from uniface.recognition import ArcFace
+        >>> recognizer = ArcFace()
+        >>> # embedding = recognizer.get_normalized_embedding(image, landmarks)
+    """
+
+    def __init__(
+        self,
+        model_name: ArcFaceWeights = ArcFaceWeights.MNET,
+        preprocessing: Optional[PreprocessConfig] = None,
+    ) -> None:
+        if preprocessing is None:
+            preprocessing = PreprocessConfig(input_mean=127.5, input_std=127.5, input_size=(112, 112))
+        model_path = verify_model_weights(model_name)
+        super().__init__(model_path=model_path, preprocessing=preprocessing)
+
+
+class MobileFace(BaseRecognizer):
+    """Lightweight MobileFaceNet model for fast face recognition.
+
+    This class provides a concrete implementation of the BaseRecognizer,
+    pre-configured for MobileFaceNet models. It is optimized for speed,
+    making it suitable for edge devices.
+
+    Args:
+        model_name (MobileFaceWeights): The specific MobileFaceNet model variant to use.
+            Defaults to `MobileFaceWeights.MNET_V2`.
+        preprocessing (Optional[PreprocessConfig]): An optional custom preprocessing
+            configuration. If None, a default config for MobileFaceNet is used.
+
+    Example:
+        >>> from uniface.recognition import MobileFace
+        >>> recognizer = MobileFace()
+        >>> # embedding = recognizer.get_normalized_embedding(image, landmarks)
+    """
+
+    def __init__(
+        self,
+        model_name: MobileFaceWeights = MobileFaceWeights.MNET_V2,
+        preprocessing: Optional[PreprocessConfig] = None,
+    ) -> None:
+        if preprocessing is None:
+            preprocessing = PreprocessConfig(input_mean=127.5, input_std=127.5, input_size=(112, 112))
+        model_path = verify_model_weights(model_name)
+        super().__init__(model_path=model_path, preprocessing=preprocessing)
+
+
+class SphereFace(BaseRecognizer):
+    """SphereFace model using angular margin for face recognition.
+
+    This class provides a concrete implementation of the BaseRecognizer,
+    pre-configured for SphereFace models, which were among the first to
+    introduce angular margin loss functions.
+
+    Args:
+        model_name (SphereFaceWeights): The specific SphereFace model variant to use.
+            Defaults to `SphereFaceWeights.SPHERE20`.
+        preprocessing (Optional[PreprocessConfig]): An optional custom preprocessing
+            configuration. If None, a default config for SphereFace is used.
+
+    Example:
+        >>> from uniface.recognition import SphereFace
+        >>> recognizer = SphereFace()
+        >>> # embedding = recognizer.get_normalized_embedding(image, landmarks)
+    """
+
+    def __init__(
+        self,
+        model_name: SphereFaceWeights = SphereFaceWeights.SPHERE20,
+        preprocessing: Optional[PreprocessConfig] = None,
+    ) -> None:
+        if preprocessing is None:
+            preprocessing = PreprocessConfig(input_mean=127.5, input_std=127.5, input_size=(112, 112))
+
+        model_path = verify_model_weights(model_name)
+        super().__init__(model_path=model_path, preprocessing=preprocessing)

uniface/retinaface.py

@@ -1,256 +0,0 @@
-# Copyright 2024 Yakhyokhuja Valikhujaev
-# Author: Yakhyokhuja Valikhujaev
-# GitHub: https://github.com/yakhyo
-
-import os
-import cv2
-import numpy as np
-import onnxruntime as ort
-
-import torch
-from typing import Tuple, List, Optional, Literal
-
-from uniface.log import Logger
-from uniface.model_store import verify_model_weights
-
-from uniface.common import (
-    nms,
-    resize_image,
-    decode_boxes,
-    generate_anchors,
-    decode_landmarks
-)
-
-
-class RetinaFace:
-    """
-    A class for face detection using the RetinaFace model.
-
-    Args:
-        model (str): Path or identifier of the model weights.
-        conf_thresh (float): Confidence threshold for detections. Defaults to 0.5.
-        nms_thresh (float): Non-maximum suppression threshold. Defaults to 0.4.
-        pre_nms_topk (int): Maximum number of detections before NMS. Defaults to 5000.
-        post_nms_topk (int): Maximum number of detections after NMS. Defaults to 750.
-        dynamic_size (Optional[bool]): Whether to adjust anchor generation dynamically based on image size. Defaults to False.
-        input_size (Optional[Tuple[int, int]]): Static input size for the model (width, height). Defaults to (640, 640).
-
-    Attributes:
-        conf_thresh (float): Confidence threshold for filtering detections.
-        nms_thresh (float): Threshold for NMS to remove duplicate detections.
-        pre_nms_topk (int): Maximum detections to consider before applying NMS.
-        post_nms_topk (int): Maximum detections retained after applying NMS.
-        dynamic_size (bool): Indicates if input size and anchors are dynamically adjusted.
-        input_size (Tuple[int, int]): The model's input image size.
-        _model_path (str): Path to the model weights.
-        _priors (torch.Tensor): Precomputed anchor boxes for static input size.
-    """
-
-    def __init__(
-        self,
-        model: str,
-        conf_thresh: float = 0.5,
-        nms_thresh: float = 0.4,
-        pre_nms_topk: int = 5000,
-        post_nms_topk: int = 750,
-        dynamic_size: Optional[bool] = False,
-        input_size: Optional[Tuple[int, int]] = (640, 640),  # Default input size if dynamic_size=False
-    ) -> None:
-
-        self.conf_thresh = conf_thresh
-        self.nms_thresh = nms_thresh
-        self.pre_nms_topk = pre_nms_topk
-        self.post_nms_topk = post_nms_topk
-        self.dynamic_size = dynamic_size
-        self.input_size = input_size
-
-        Logger.info(
-            f"Initializing RetinaFace with model={model}, conf_thresh={conf_thresh}, nms_thresh={nms_thresh}, "
-            f"pre_nms_topk={pre_nms_topk}, post_nms_topk={post_nms_topk}, dynamic_size={dynamic_size}, "
-            f"input_size={input_size}"
-        )
-
-        # Get path to model weights
-        self._model_path = verify_model_weights(model)
-        Logger.info(f"Verified model weights located at: {self._model_path}")
-
-        # Precompute anchors if using static size
-        if not dynamic_size and input_size is not None:
-            self._priors = generate_anchors(image_size=input_size)
-            Logger.debug("Generated anchors for static input size.")
-
-        # Initialize model
-        self._initialize_model(self._model_path)
-
-    def _initialize_model(self, model_path: str) -> None:
-        """
-        Initializes an ONNX model session from the given path.
-
-        Args:
-            model_path (str): The file path to the ONNX model.
-
-        Raises:
-            RuntimeError: If the model fails to load, logs an error and raises an exception.
-        """
-        try:
-            self.session = ort.InferenceSession(model_path)
-            self.input_name = self.session.get_inputs()[0].name
-            Logger.info(f"Successfully initialized the model from {model_path}")
-        except Exception as e:
-            Logger.error(f"Failed to load model from '{model_path}': {e}")
-            raise RuntimeError(f"Failed to initialize model session for '{model_path}'") from e
-
-    def preprocess(self, image: np.ndarray) -> np.ndarray:
-        """Preprocess input image for model inference.
-
-        Args:
-            image (np.ndarray): Input image.
-
-        Returns:
-            np.ndarray: Preprocessed image tensor with shape (1, C, H, W)
-        """
-        image = np.float32(image) - np.array([104, 117, 123], dtype=np.float32)
-        image = image.transpose(2, 0, 1)  # HWC to CHW
-        image = np.expand_dims(image, axis=0)  # Add batch dimension (1, C, H, W)
-        return image
-
-    def inference(self, input_tensor: np.ndarray) -> List[np.ndarray]:
-        """Perform model inference on the preprocessed image tensor.
-
-        Args:
-            input_tensor (np.ndarray): Preprocessed input tensor.
-
-        Returns:
-            Tuple[np.ndarray, np.ndarray]: Raw model outputs.
-        """
-        return self.session.run(None, {self.input_name: input_tensor})
-
-    def detect(
-        self,
-        image: np.ndarray,
-        max_num: Optional[int] = 0,
-        metric: Literal["default", "max"] = "default",
-        center_weight: Optional[float] = 2.0
-    ) -> Tuple[np.ndarray, np.ndarray]:
-        """
-        Perform face detection on an input image and return bounding boxes and landmarks.
-
-        Args:
-            image (np.ndarray): Input image as a NumPy array of shape (height, width, channels).
-            max_num (int, optional): Maximum number of detections to return. Defaults to 1.
-            metric (str, optional): Metric for ranking detections when `max_num` is specified. 
-                Options:
-                - "default": Prioritize detections closer to the image center.
-                - "max": Prioritize detections with larger bounding box areas.
-            center_weight (float, optional): Weight for penalizing detections farther from the image center 
-                when using the "default" metric. Defaults to 2.0.
-
-        Returns:
-            Tuple[np.ndarray, np.ndarray]: Detection results containing:
-                - detections (np.ndarray): Array of detected bounding boxes with confidence scores.
-                Shape: (num_detections, 5), where each row is [x_min, y_min, x_max, y_max, score].
-                - landmarks (np.ndarray): Array of detected facial landmarks.
-                Shape: (num_detections, 5, 2), where each row contains 5 landmark points (x, y).
-        """
-
-        if self.dynamic_size:
-            height, width, _ = image.shape
-            self._priors = generate_anchors(image_size=(height, width))  # generate anchors for each input image
-            resize_factor = 1.0  # No resizing
-        else:
-            image, resize_factor = resize_image(image, target_shape=self.input_size)
-
-        height, width, _ = image.shape
-        image_tensor = self.preprocess(image)
-
-        # ONNXRuntime inference
-        outputs = self.inference(image_tensor)
-
-        # Postprocessing
-        detections, landmarks = self.postprocess(outputs, resize_factor, shape=(width, height))
-
-        if max_num > 0 and detections.shape[0] > max_num:
-            # Calculate area of detections
-            areas = (detections[:, 2] - detections[:, 0]) * (detections[:, 3] - detections[:, 1])
-
-            # Calculate offsets from image center
-            center = (height // 2, width // 2)
-            offsets = np.vstack([
-                (detections[:, 0] + detections[:, 2]) / 2 - center[1],
-                (detections[:, 1] + detections[:, 3]) / 2 - center[0]
-            ])
-            offset_dist_squared = np.sum(np.power(offsets, 2.0), axis=0)
-
-            # Calculate scores based on the chosen metric
-            if metric == 'max':
-                scores = areas
-            else:
-                scores = areas - offset_dist_squared * center_weight
-
-            # Sort by scores and select top `max_num`
-            sorted_indices = np.argsort(scores)[::-1][:max_num]
-
-            detections = detections[sorted_indices]
-            landmarks = landmarks[sorted_indices]
-
-        return detections, landmarks
-
-    def postprocess(self, outputs: List[np.ndarray], resize_factor: float, shape: Tuple[int, int]) -> Tuple[np.ndarray, np.ndarray]:
-        """
-        Process the model outputs into final detection results.
-
-        Args:
-            outputs (List[np.ndarray]): Raw outputs from the detection model.
-                - outputs[0]: Location predictions (bounding box coordinates).
-                - outputs[1]: Class confidence scores.
-                - outputs[2]: Landmark predictions.
-            resize_factor (float): Factor used to resize the input image during preprocessing.
-            shape (Tuple[int, int]): Original shape of the image as (height, width).
-
-        Returns:
-            Tuple[np.ndarray, np.ndarray]: Processed results containing:
-                - detections (np.ndarray): Array of detected bounding boxes with confidence scores.
-                Shape: (num_detections, 5), where each row is [x_min, y_min, x_max, y_max, score].
-                - landmarks (np.ndarray): Array of detected facial landmarks.
-                Shape: (num_detections, 5, 2), where each row contains 5 landmark points (x, y).
-        """
-        loc, conf, landmarks = outputs[0].squeeze(0), outputs[1].squeeze(0), outputs[2].squeeze(0)
-
-        # Decode boxes and landmarks
-        boxes = decode_boxes(torch.tensor(loc), self._priors).cpu().numpy()
-        landmarks = decode_landmarks(torch.tensor(landmarks), self._priors).cpu().numpy()
-
-        boxes, landmarks = self._scale_detections(boxes, landmarks, resize_factor, shape=(shape[0], shape[1]))
-
-        # Extract confidence scores for the face class
-        scores = conf[:, 1]
-        mask = scores > self.conf_thresh
-
-        # Filter by confidence threshold
-        boxes, landmarks, scores = boxes[mask], landmarks[mask], scores[mask]
-
-        # Sort by scores
-        order = scores.argsort()[::-1][:self.pre_nms_topk]
-        boxes, landmarks, scores = boxes[order], landmarks[order], scores[order]
-
-        # Apply NMS
-        detections = np.hstack((boxes, scores[:, np.newaxis])).astype(np.float32, copy=False)
-        keep = nms(detections, self.nms_thresh)
-        detections, landmarks = detections[keep], landmarks[keep]
-
-        # Keep top-k detections
-        detections, landmarks = detections[:self.post_nms_topk], landmarks[:self.post_nms_topk]
-
-        landmarks = landmarks.reshape(-1, 5, 2).astype(np.int32)
-
-        return detections, landmarks
-
-    def _scale_detections(self, boxes: np.ndarray, landmarks: np.ndarray, resize_factor: float, shape: Tuple[int, int]) -> Tuple[np.ndarray, np.ndarray]:
-        """Scale bounding boxes and landmarks to the original image size."""
-        bbox_scale = np.array([shape[0], shape[1]] * 2)
-        boxes = boxes * bbox_scale / resize_factor
-
-        landmark_scale = np.array([shape[0], shape[1]] * 5)
-        landmarks = landmarks * landmark_scale / resize_factor
-
-        return boxes, landmarks

uniface/version.py

@@ -1,15 +0,0 @@
-# Copyright 2024 Yakhyokhuja Valikhujaev
-#
-# Licensed under the MIT License.
-# You may obtain a copy of the License at
-#
-#     https://opensource.org/licenses/MIT
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-__version__ = "0.1.1"
-__author__ = "Yakhyokhuja Valikhujaev"

uniface/visualization.py

@@ -1,38 +1,330 @@
-# Copyright 2024 Yakhyokhuja Valikhujaev
+# Copyright 2025 Yakhyokhuja Valikhujaev
 # Author: Yakhyokhuja Valikhujaev
 # GitHub: https://github.com/yakhyo
 
+from typing import List, Tuple, Union
+
 import cv2
 import numpy as np
 
+# Face parsing component names (19 classes)
+FACE_PARSING_LABELS = [
+    'background',
+    'skin',
+    'l_brow',
+    'r_brow',
+    'l_eye',
+    'r_eye',
+    'eye_g',
+    'l_ear',
+    'r_ear',
+    'ear_r',
+    'nose',
+    'mouth',
+    'u_lip',
+    'l_lip',
+    'neck',
+    'neck_l',
+    'cloth',
+    'hair',
+    'hat',
+]
 
-def draw_detections(image, detections, vis_threshold=0.6):
+# Color palette for face parsing visualization
+FACE_PARSING_COLORS = [
+    [0, 0, 0],
+    [255, 85, 0],
+    [255, 170, 0],
+    [255, 0, 85],
+    [255, 0, 170],
+    [0, 255, 0],
+    [85, 255, 0],
+    [170, 255, 0],
+    [0, 255, 85],
+    [0, 255, 170],
+    [0, 0, 255],
+    [85, 0, 255],
+    [170, 0, 255],
+    [0, 85, 255],
+    [0, 170, 255],
+    [255, 255, 0],
+    [255, 255, 85],
+    [255, 255, 170],
+    [255, 0, 255],
+]
+
+
+def draw_detections(
+    *,
+    image: np.ndarray,
+    bboxes: Union[List[np.ndarray], List[List[float]]],
+    scores: Union[np.ndarray, List[float]],
+    landmarks: Union[List[np.ndarray], List[List[List[float]]]],
+    vis_threshold: float = 0.6,
+    draw_score: bool = False,
+    fancy_bbox: bool = True,
+):
     """
-    Draw bounding boxes and landmarks on the image.
+    Draws bounding boxes, landmarks, and optional scores on an image.
 
     Args:
-        image (ndarray): Image to draw detections on.
-        detections (tuple): (bounding boxes, landmarks) as NumPy arrays.
-        vis_threshold (float): Confidence threshold for filtering detections.
+        image: Input image to draw on.
+        bboxes: List of bounding boxes [x1, y1, x2, y2].
+        scores: List of confidence scores.
+        landmarks: List of landmark sets with shape (5, 2).
+        vis_threshold: Confidence threshold for filtering. Defaults to 0.6.
+        draw_score: Whether to draw confidence scores. Defaults to False.
     """
+    colors = [(0, 0, 255), (0, 255, 255), (255, 0, 255), (0, 255, 0), (255, 0, 0)]
 
-    _colors = [(0, 0, 255), (0, 255, 255), (255, 0, 255), (0, 255, 0), (255, 0, 0)]
-
-    # Unpack detections
-    boxes, landmarks = detections
-    scores = boxes[:, 4]
+    # Calculate line thickness based on image size
+    line_thickness = max(round(sum(image.shape[:2]) / 2 * 0.003), 2)
 
     # Filter detections by confidence threshold
-    filtered = scores >= vis_threshold
-    boxes = boxes[filtered, :4].astype(np.int32)
-    landmarks = landmarks[filtered]
-    scores = scores[filtered]
+    keep_indices = [i for i, score in enumerate(scores) if score >= vis_threshold]
 
-    print(f"#faces: {len(scores)}")
+    for i in keep_indices:
+        bbox = np.array(bboxes[i], dtype=np.int32)
+        score = scores[i]
+        landmark_set = np.array(landmarks[i], dtype=np.int32)
 
-    # Draw bounding boxes, scores, and landmarks
-    for box, score, landmark in zip(boxes, scores, landmarks):
-        cv2.rectangle(image, box[:2], box[2:], (0, 0, 255), 2)
-        cv2.putText(image, f"{score:.2f}", (box[0], box[1] + 12), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 255), 1)
-        for point, color in zip(landmark, _colors):
-            cv2.circle(image, tuple(point), 2, color, -1)
+        # Calculate dynamic font scale based on bbox height
+        bbox_h = bbox[3] - bbox[1]
+        font_scale = max(0.4, min(0.7, bbox_h / 200))
+        font_thickness = 2
+
+        # Draw bounding box
+        if fancy_bbox:
+            draw_fancy_bbox(image, bbox, color=(0, 255, 0), thickness=line_thickness, proportion=0.2)
+        else:
+            cv2.rectangle(image, tuple(bbox[:2]), tuple(bbox[2:]), (0, 255, 0), line_thickness)
+
+        # Draw confidence score with background
+        if draw_score:
+            text = f'{score:.2f}'
+            (text_width, text_height), baseline = cv2.getTextSize(
+                text, cv2.FONT_HERSHEY_SIMPLEX, font_scale, font_thickness
+            )
+
+            # Draw background rectangle
+            cv2.rectangle(
+                image,
+                (bbox[0], bbox[1] - text_height - baseline - 10),
+                (bbox[0] + text_width + 10, bbox[1]),
+                (0, 255, 0),
+                -1,
+            )
+
+            # Draw text
+            cv2.putText(
+                image,
+                text,
+                (bbox[0] + 5, bbox[1] - 5),
+                cv2.FONT_HERSHEY_SIMPLEX,
+                font_scale,
+                (0, 0, 0),
+                font_thickness,
+            )
+
+        # Draw landmarks
+        for j, point in enumerate(landmark_set):
+            cv2.circle(image, tuple(point), line_thickness + 1, colors[j], -1)
+
+
+def draw_fancy_bbox(
+    image: np.ndarray,
+    bbox: np.ndarray,
+    color: Tuple[int, int, int] = (0, 255, 0),
+    thickness: int = 3,
+    proportion: float = 0.2,
+):
+    """
+    Draws a bounding box with fancy corners on an image.
+
+    Args:
+        image: Input image to draw on.
+        bbox: Bounding box coordinates [x1, y1, x2, y2].
+        color: Color of the bounding box. Defaults to green.
+        thickness: Thickness of the bounding box lines. Defaults to 3.
+        proportion: Proportion of the corner length to the width/height of the bounding box. Defaults to 0.2.
+    """
+    x1, y1, x2, y2 = map(int, bbox)
+    width = x2 - x1
+    height = y2 - y1
+
+    corner_length = int(proportion * min(width, height))
+
+    # Draw the rectangle
+    cv2.rectangle(image, (x1, y1), (x2, y2), color, 1)
+
+    # Top-left corner
+    cv2.line(image, (x1, y1), (x1 + corner_length, y1), color, thickness)
+    cv2.line(image, (x1, y1), (x1, y1 + corner_length), color, thickness)
+
+    # Top-right corner
+    cv2.line(image, (x2, y1), (x2 - corner_length, y1), color, thickness)
+    cv2.line(image, (x2, y1), (x2, y1 + corner_length), color, thickness)
+
+    # Bottom-left corner
+    cv2.line(image, (x1, y2), (x1, y2 - corner_length), color, thickness)
+    cv2.line(image, (x1, y2), (x1 + corner_length, y2), color, thickness)
+
+    # Bottom-right corner
+    cv2.line(image, (x2, y2), (x2, y2 - corner_length), color, thickness)
+    cv2.line(image, (x2, y2), (x2 - corner_length, y2), color, thickness)
+
+
+def draw_gaze(
+    image: np.ndarray,
+    bbox: np.ndarray,
+    pitch: np.ndarray,
+    yaw: np.ndarray,
+    *,
+    draw_bbox: bool = True,
+    fancy_bbox: bool = True,
+    draw_angles: bool = True,
+):
+    """
+    Draws gaze direction with optional bounding box on an image.
+
+    Args:
+        image: Input image to draw on (modified in-place).
+        bbox: Face bounding box [x1, y1, x2, y2].
+        pitch: Vertical gaze angle in radians.
+        yaw: Horizontal gaze angle in radians.
+        draw_bbox: Whether to draw the bounding box. Defaults to True.
+        fancy_bbox: Use fancy corner-style bbox. Defaults to True.
+        draw_angles: Whether to display pitch/yaw values as text. Defaults to False.
+    """
+    x_min, y_min, x_max, y_max = map(int, bbox[:4])
+
+    # Calculate dynamic line thickness based on image size (same as draw_detections)
+    line_thickness = max(round(sum(image.shape[:2]) / 2 * 0.003), 2)
+
+    # Calculate dynamic font scale based on bbox height (same as draw_detections)
+    bbox_h = y_max - y_min
+    font_scale = max(0.4, min(0.7, bbox_h / 200))
+    font_thickness = 2
+
+    # Draw bounding box if requested
+    if draw_bbox:
+        if fancy_bbox:
+            draw_fancy_bbox(image, bbox, color=(0, 255, 0), thickness=line_thickness)
+        else:
+            cv2.rectangle(image, (x_min, y_min), (x_max, y_max), (0, 255, 0), line_thickness)
+
+    # Calculate center of the bounding box
+    x_center = (x_min + x_max) // 2
+    y_center = (y_min + y_max) // 2
+
+    # Calculate the direction of the gaze
+    length = x_max - x_min
+    dx = int(-length * np.sin(pitch) * np.cos(yaw))
+    dy = int(-length * np.sin(yaw))
+
+    point1 = (x_center, y_center)
+    point2 = (x_center + dx, y_center + dy)
+
+    # Calculate dynamic center point radius based on line thickness
+    center_radius = max(line_thickness + 1, 4)
+
+    # Draw gaze direction
+    cv2.circle(image, (x_center, y_center), radius=center_radius, color=(0, 0, 255), thickness=-1)
+    cv2.arrowedLine(
+        image,
+        point1,
+        point2,
+        color=(0, 0, 255),
+        thickness=line_thickness,
+        line_type=cv2.LINE_AA,
+        tipLength=0.25,
+    )
+
+    # Draw angle values
+    if draw_angles:
+        text = f'P:{np.degrees(pitch):.0f}deg Y:{np.degrees(yaw):.0f}deg'
+        (text_width, text_height), baseline = cv2.getTextSize(
+            text, cv2.FONT_HERSHEY_SIMPLEX, font_scale, font_thickness
+        )
+
+        # Draw background rectangle for text
+        cv2.rectangle(
+            image,
+            (x_min, y_min - text_height - baseline - 10),
+            (x_min + text_width + 10, y_min),
+            (0, 0, 255),
+            -1,
+        )
+
+        # Draw text
+        cv2.putText(
+            image,
+            text,
+            (x_min + 5, y_min - 5),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            font_scale,
+            (255, 255, 255),
+            font_thickness,
+        )
+
+
+def vis_parsing_maps(
+    image: np.ndarray,
+    segmentation_mask: np.ndarray,
+    *,
+    save_image: bool = False,
+    save_path: str = 'result.png',
+) -> np.ndarray:
+    """
+    Visualizes face parsing segmentation mask by overlaying colored regions on the image.
+
+    Args:
+        image: Input face image in RGB format with shape (H, W, 3).
+        segmentation_mask: Segmentation mask with shape (H, W) where each pixel
+                          value represents a facial component class (0-18).
+        save_image: Whether to save the visualization to disk. Defaults to False.
+        save_path: Path to save the visualization if save_image is True.
+
+    Returns:
+        np.ndarray: Blended image with segmentation overlay in BGR format.
+
+    Example:
+        >>> import cv2
+        >>> from uniface.parsing import BiSeNet
+        >>> from uniface.visualization import vis_parsing_maps
+        >>>
+        >>> parser = BiSeNet()
+        >>> face_image = cv2.imread('face.jpg')
+        >>> mask = parser.parse(face_image)
+        >>>
+        >>> # Visualize
+        >>> face_rgb = cv2.cvtColor(face_image, cv2.COLOR_BGR2RGB)
+        >>> result = vis_parsing_maps(face_rgb, mask)
+        >>> cv2.imwrite('parsed_face.jpg', result)
+    """
+    # Create numpy arrays for image and segmentation mask
+    image = np.array(image).copy().astype(np.uint8)
+    segmentation_mask = segmentation_mask.copy().astype(np.uint8)
+
+    # Create a color mask
+    segmentation_mask_color = np.zeros((segmentation_mask.shape[0], segmentation_mask.shape[1], 3))
+
+    num_classes = np.max(segmentation_mask)
+
+    for class_index in range(1, num_classes + 1):
+        class_pixels = np.where(segmentation_mask == class_index)
+        segmentation_mask_color[class_pixels[0], class_pixels[1], :] = FACE_PARSING_COLORS[class_index]
+
+    segmentation_mask_color = segmentation_mask_color.astype(np.uint8)
+
+    # Convert image to BGR format for blending
+    bgr_image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
+
+    # Blend the image with the segmentation mask
+    blended_image = cv2.addWeighted(bgr_image, 0.6, segmentation_mask_color, 0.4, 0)
+
+    # Save the result if required
+    if save_image:
+        cv2.imwrite(save_path, blended_image, [int(cv2.IMWRITE_JPEG_QUALITY), 100])
+
+    return blended_image