An Experimental Implementation of Face Verification, 96.8% on LFW.
A messy code for developing a face verfication program.
It includes a C++ face detection / alignment program, joint bayesian and several supplementary codes. My Caffe model define file is also provided. Note that I use a fresh layer called Insanity to replace the ReLU activation. The Insanity layer can be found in my Caffe repository. Please feel free to use the codes if you need.
If you are also interested in face verification, please contact me via the issue.
The CASIA-webface dataset is really very dirty, and I believe that if someone could wash it up, the accuracy would increase further. If you did so, please kindly contact me. I will pay for it.
Good News: @潘泳苹果皮 and his colleagues have washed the CASIA-webface database manually. After washing, 27703 wrong images are deleted. The washed list can be downloaded from http://pan.baidu.com/s/1hrKpbm8 . Great thanks to them!
2017/04/21 The project page of my new paper, NormFace: L2 HyperSphere Embedding for Face Verification, is created on https://github.com/happynear/NormFace. Using the new loss functions to fine-tune a network, the accuracy will increase a little bit higher.
2017/02/18 I trained a center-face model on MS-Celeb-1M dataset and get 99.3% on LFW. Here is the model (http://pan.baidu.com/s/1jIJT4Rc) and the aligned LFW images (http://pan.baidu.com/s/1bp7qzJh). To run the evaluation, you need to load LFW pairs by
aligned_lfw.zip, extract feature by
ReadFeatureLFW.mand get the accuracy by
lfwPCA.m. The function
lfwPCA.mis from pdollar-toolbox.
Recently I talked with Yandong Wen. He said that there were more than 1,000 identities' overlap between MS-Celeb-1M and LFW. So this accuracy, 99.3%, is not a reliable value, and the performance on other datasets may not be very good.
2015/07/05 Added a matlab face alignment wrapper (MatAlignment.cpp). Now you can do the face alignment job in matlab. A demo (VerificationDemo.m) is also privided.
Training DeepID (pure softmax network).
create database. done.
iteration 360,000, lr=0.01,
lfw verification: L2 : 95.9%, jb :
iteration 500,000, lr=0.001,
lfw verification: L2 : 96.8%, jb : 93.3% (strongly overfit, it's >99% for lfw training set).
iteration 660,000, lr=0.0001,
lfw verification: L2 : 96.78% (converged)
Accuracy on training set is about 89.5%~91.5%. LFW result with L2 or cosine has reached what the paper claimed. Joint Bayesian seems to be strongly overfit. The main reason is that I only train Joint Bayesian on the lfw training set, not CASIA-WebFace. Joint Bayesian for over 10,000 classes is too costy for my PC.
This model is public available now: http://pan.baidu.com/s/1qXhNOZE . Mean file is in http://pan.baidu.com/s/1eQYEJU6 .
Another model with resolution of 64*64 is trained. By ensembling the two models, accuracy increases to 97.18%.
Training DeepID2 (siamese network)
create database. done.
My GPU memory is only 2GB. It is too slow to train a siamese network. I need Titan X!!!