insightface/iccv19-challenge/README.md

The Lightweight Face Recognition Challenge & Workshop will be held in conjunction with the International Conference on Computer Vision (ICCV) 2019, Seoul Korea.

Test Server

==================

How To Start:

Training:

1. Download ms1m-retinaface from baiducloud or dropbox and unzip it to $INSIGHTFACE_ROOT/datasets/
2. Go into $INSIGHTFACE_ROOT/recognition/
3. Refer to the retina dataset config section in sample_config.py and copy it to your ownconfig.py.
4. Start training with CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --dataset retina --network [your-network] --loss arcface. It will output the accuracy of lfw, cfp_fp and agedb_30 every 2000 batches by default.

---

Testing:

1. testdata-image from baiducloud or dropbox. These face images are all pre-processed and aligned so no need to do further modification.
2. To download testdata-video from iQIYI, please visit http://challenge.ai.iqiyi.com/data-cluster. You must download iQIYI-VID-FACE.z01, iQIYI-VID-FACE.z02 and iQIYI-VID-FACE.zip after signin. These face images are all pre-processed and aligned so no need to do further modification.
   1. To unzip: zip iQIYI_VID_FACE.zip -s=0 --out iQIYI_VID_FACE_ALL.zip; unzip iQIYI_VID_FACE_ALL.zip
   2. We can get a directory named iQIYI_VID_FACE after decompression. Then we have to move video_filelist.txt in testdata-image package to iQIYI_VID_FACE/filelist.txt, to indicate the order of videos in our submission feature file.
3. To generate image feature submission file: check gen_image_feature.py
4. To generate video feature submission file: check gen_video_feature.py
5. Submit binary feature to the right section on test server.

You can also check the verification performance during training time on LFW,CFP_FP,AgeDB_30 datasets.

---

Evaluation:

Final ranking is determined by the TAR under 1:1 protocal only, for all valid submissions.

For image testset, we evaluate the TAR under FAR@e-8 while we choose the TAR under FAR@e-4 for video testset.

For track-1, we will rank all players as following formula: TAR(glint-light)+TAR(iqiyi-light)

For track-2, we will rank all players as following formula: TAR(glint-large)+TAR(iqiyi-large)

---

Discussion:

https://github.com/deepinsight/insightface/issues/632

---

Baseline:

1. Network y2(a deeper mobilefacenet): 933M FLOPs. TAR_image: 0.64691, TAR_video: [TODO]
2. Network r100fc(ResNet100FC-IR): 24G FLOPs. TAR_image: 0.80312, TAR_video: [TODO]

---

Candidate solutions:

1. Use slightly deeper or wider mobile-level networks.
2. Try different training methods/losses than straightforward arcface.
3. OctConv, to reduce FLOPs.
4. HRNet, for large FLOPs track. and so on