[INFO] A better CNN pre-trained model

[joyhuang9473]

author: AlfredXiangWu

reference: https://github.com/AlfredXiangWu/face_verification_experiment

related: #5

• caffe framework
• CASIA-WebFace dataset: contains 493456 face images of 10575 identities.
• feature extraction code: https://github.com/AlfredXiangWu/python_misc/blob/master/caffe/caffe_ftr.py
• accuracy: (The single convolution net testing is evaluated on unsupervised setting only computing cosine similarity for lfw pairs. )
  • Dataset A: 97.77%
  • Dataset B: 98.13%

[joyhuang9473]

author: happynear

reference: https://github.com/happynear/FaceVerification

DeepID

• caffemodel: http://pan.baidu.com/s/1qXhNOZE
• deploy file: https://github.com/happynear/FaceVerification/blob/master/caffe_proto/CASIA_deploy.prototxt
• mean file: http://pan.baidu.com/s/1eQYEJU6
• lfw verification: L2 : 96.78% (converged)

LFW result with L2 or cosine has reached what the paper claimed.

Another model with resolution of 64*64 is trained. By ensembling the two models, accuracy increases to 97.18%.

I used all the database to train the model. I didn't split it into train and val subset.

lfwL2.m

[joyhuang9473]

author: cmusatyalab

reference: https://github.com/cmusatyalab/openface

Face recognition with Google's FaceNet deep neural network.

• dataset: CASIA-Webface + FaceScrub

note:

openface/openface/align_dlib.py

[joyhuang9473]

author: davidsandberg

reference: https://github.com/davidsandberg/facenet/tree/master/facenet

TensorFlow implementation, Google's FaceNet

• dataset: CASIA-Webface + FaceScrub
• pre-trained model: https://drive.google.com/file/d/0B5MzpY9kBtDVUFNFR21kZnYybkE/view?usp=sharing
• lfw verification: 0.916±0.010 ("model-20160306.ckpt-500000")

Pre-processing: The data has been pre-processed as described on the OpenFace web page (https://cmusatyalab.github.io/openface/models-and-accuracies/), i.e. using ./util/align-dlib.py data/lfw/raw align outerEyesAndNose data/lfw/dlib-affine-sz:96 --size 96 --fallbackLfw data/lfw/deepfunneled

Performance: The accuracy on LFW for the model "model-20160306.ckpt-500000" is 0.916±0.010. The test can be run using "validate_on_lfw.py".