Face Classification: A Specialized Benchmark Study

Code for face classification benchmark of the following paper

Face Classification: A Specialzed Benchmark Study

Jiali Duan, Shengcai Liao, Shuai Zhou, Stan Z. Li To appear in CCBR 2016 oral

Requirements

0. A recent installation of caffe with its python wrapper.
1. All other python requirements are mentioned in requirements.txt
2. Matlab R2013a or higher
3. py-faster-rcnn repository
4. WIDER FACE benchmark

Getting started

Face Classification Benchmark

The proposed face classification benchmark contains more than 3.5 million extracted patches, with 198616 face patches in pos-new.tgz and 3359526 nonface patches in directory neg-new2/ .

Nonface patches are compressed into 18 seperate parts, download them and run the following command:

cd neg-new2/
cat face_nonface_patches_part* > face_nonface_patches.tgz
tar -zxvf face_nonface_patches.tgz

RPN network

Pretrained models for extracting proposals are tarred into Pretrained_model.tgz with zf_faster_rcnn_iter_130000.caffemodel, train.prototxt, test.prototxt and solver.prototxt included. We have made some modifications on the original py-faster-rcnn code to train a specialized faster-rcnn network for proposal extraction.

Replace the original py-faster-rcnn directories with corresponding directories provided in Proposal_Extraction_Code. Run the face_proposal.py to extract proposals.

Make sure to change the following variables in the file according to your developing environment:

0. save_dir is the path to save proposals
1. data_root is the parent folder of JPEGImages of WIDER FACE
2. flabel is the path to the label file. We have provided our training and testing list files in train_test_list.tgz
3. Add faster-rcnn-caffe-python path to your PYTHONPATH using the following command:

export PYTHONPATH='/path/to/you/caffe-fast-rcnn/python':$PYTHONPATH

You should get similar results saved under pos and neg directories of save_dir:

Evaluation

All the neccessary codes and scores with respect to each basline algorithm are included in Evaluation_Code directory. You can simply run ShowROCCurve.m to get the final ROC curce. You should be able to get the following result.

Baseline

Training and Testing Code for each baseline algorithm would be released

lomo+boosting

Evaluate: files needed to test lomo+boosting can be downloaded via http://pan.baidu.com/s/1o7SlPqi and lomo+boost/Testing including test set of positive and negative samples(4D_Face_test.mat, 4D_Nonface_test.mat). You could extract corresponding LOMO features yourself or use extracted features saved in feat_test.mat. Final lomo+boosting model is also available：lomo_boost.mat.

LOMO+Boosting score can be obtained through:

load('lomo_boost.mat','model');
load feat_test.mat;
save_dir='lomo_score.txt'
[Fx,~]=TestGAB(model,feat_test);
dlmwrite(save_dir,Fx);

Train: inflate the directory lomo+boost/Training/data with data downloaded from http://pan.baidu.com/s/1o8xZSKE. Enter lomo+boost/Training/src run TrainDetector_Demo.m and the training will be on the fly. Final model would be saved in result.mat, and you can get the inference model by extracting its model parameter.

CNN_Cifar

Evaluate: enter into CNN_Cifar/Testing and run python Evaluate.py to get the score. Make sure that you change data_root and caffe_root according to your environement.

Train: files needed to train cifar10 based CNN are included in CNN_Cifar/Training. Training LMDB is available at http://pan.baidu.com/s/1c29P6KC. Run ./train_quick.sh to train the CNN_Cifar model.

Work in progress

I'm in process of cleaning up my original evaluation code.

TODO:

• LBP training and evaluation code.
• MBLBP training and evaluation code.
• NPD training and evaluation code.
• Cascade CNN baseline.