This is an implementation of feature extraction method with center loss function and the small network SqueezeNet. The network is trained on CASIA_WebFace with preprocessing of face cropping and affine transform.
We tested the verification accuracy on LFW dataset and got an accuracy of 97.6%.
After training, if you truncate the fully connected layers in SqueezeNet, the caffemodel will be only 2.6 MB.
How to train the model
1, download and make the caffe with center loss (see the link in the paper)
2, download a proper face dataset(better with cropped face), resize the images to the same size. (I used 226X264)
3, create the anno.txt with each face labelled, see the format in data/anno.txt
4, run the python script (This script is written in python 2.7)
The loss change during training is:
center loss increased at the beginning, but decreased a little approaching the end
softmax loss decreased all the way
the total loss decreased all the way (ratio between center loss and softmax loss is 0.008:1)
