Home

Awesome

Sphereface-model

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%.

I think the performance can still be further improved by carefully fine-tuning. Feel free to use this model.

Experiment Results

Train 22000(lambda=3.6, batch_size=170).

The distribution of features on LFW:

Feature distribution

The roc curve:

ROC

Accuracy on LFW:

OriginalWith PCAWith mirror trickWith 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].

Update

Reference

https://github.com/wy1iu/sphereface

https://github.com/happynear/NormFace/blob/master/MirrorFace.md