This is the pre-trained model of SphereFace : Deep Hypersphere Embedding for Face Recognition. This model is trained on CASIA-Webface and the accuracy on LFW is 99.18%.
- Training on Webface with the default setting only got an accuracy of 98.5%， which is much lower than the paper claimed(~99.26%).
- Fixed lambda = 5, kept on training, and got an accuracy of 98.8%.
- Decreased the lambda. Final accuracy is 99.18%.
I think the performance can still be further improved by carefully fine-tuning. Feel free to use this model.
Train 22000(lambda=3.6, batch_size=170).
The distribution of features on LFW:
The roc curve:
Accuracy on LFW:
| Original | With PCA | With mirror trick | With mirror trick and PCA |
|---|---|---|---|
| 98.88% | 99.13% | 98.98% | 99.18% |
Here is the model: https://pan.baidu.com/s/1pL0pmll
LFW evaluation code can be found in [https://github.com/happynear/FaceVerification].
- Accuracy on LFW 99.18% 2017.07.31
https://github.com/wy1iu/sphereface
https://github.com/happynear/NormFace/blob/master/MirrorFace.md