uptodiff / CosFace

Tensorflow implementation for paper CosFace: Large Margin Cosine Loss for Deep Face Recognition

Recent Update

2018.07.04: I achieved a better accuracy(99.2%) on LFW. I did some modification as bellow:

• Align webface and lfw dataset to 112x112 using insightface align method
• Set a bigger margin parameter (0.35) and a higher feature embedding demension (1024)
• Use the clean dataset and the details can be seen this

CosFace

This project is aimmed at implementing the CosFace described by the paper CosFace: Large Margin Cosine Loss for Deep Face Recognition. The code can be trained on CASIA-Webface and the best accuracy LFW is 98.6%. The result is lower than reported by paper(99.33%), which may be caused by sphere network implemented in tensorflow. I train the sphere network implemented in tensorflow using the softmax loss and just obtain the accuracy 95.6%, which is more lower than caffe version(97.88%).

Preprocessing

I supply the preprocessed dataset in baidu pan:CASIA-WebFace-112X96,lfw-112X96. You can download and unzip them to dir dataset.

If you want to preprocess the dataset by yourself, you can refer to sphereface.

Train

./train.sh

Test

Modify the MODEL_DIR in test.sh and run ./test/sh.

If you do not want to train your model, you can download my trained model and unzip it to models dir.

Reference

• facenet