ZF/insightface
1
0
Fork 0
mirror of https://github.com/deepinsight/insightface.git synced 2025-12-30 08:02:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update README.md

Browse Source
This commit is contained in:
JiankangDeng
2020-08-10 13:09:42 +01:00
committed by GitHub
parent 04e640cba3
commit 0a08c7a4e9
1 changed files with 9 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
recognition/SubCenter-ArcFace/README.md
View File

@@ -1,19 +1,24 @@
## Subcenter ArcFace
### 1. Main Contribution
### 1. Motivation
The training process of Subcenter ArcFace is almost same as [ArcFace](https://github.com/deepinsight/insightface/tree/master/recognition/ArcFace), except for one extra hyperparameter (subcenter number `loss_K`) to relax the intra-class compactness constraint. In our experiments, we find ``loss_K=3`` can achieve a good balance between accuracy and robustness.
We introduce one extra hyperparameter (subcenter number `loss_K`) to ArcFace to relax the intra-class compactness constraint. In our experiments, we find ``loss_K=3`` can achieve a good balance between accuracy and robustness.
![difference](https://github.com/deepinsight/insightface/blob/master/resources/subcenterarcfacediff.png)
### 2. Implementation
The training process of Subcenter ArcFace is almost same as [ArcFace](https://github.com/deepinsight/insightface/tree/master/recognition/ArcFace)
The increased GPU memory consumption can be easily alleviated by our parallel framework.
![framework](https://github.com/deepinsight/insightface/blob/master/resources/subcenterarcfaceframework.png)
### 2. Training Dataset
### 3. Training Dataset
1. MS1MV0 (The noise rate is around 50%), download link ([baidulink](https://pan.baidu.com/s/1bSamN5CLiSrxOuGi-Lx7tw), code ``8ql0``) ([googledrive](TODO))
### 3. Training Steps
### 4. Training Steps
1). Train Sub-center ArcFace (``loss_K=3``) on MS1MV0.