Regularizing Face Verification Nets For Pain Intensity Regression

https://arxiv.org/abs/1702.06925

1. My Caffe (https://github.com/happynear/caffe-windows/tree/ms). If you don't want to train with class-balance sampling (image_data_layer.cpp) and observing Pearson Correlation during training (correlation_loss_layer.cpp), you may use the official Caffe.
2. Matlab,
3. GPU with CUDA support,
4. MTCNN face and facial landmark detector(https://github.com/kpzhang93/MTCNN_face_detection_alignment).

We uploaded some of the predicted values generated by our algorithm in /results. You may load them in Matlab and run scripts/get_accuracy.m to get the performance under various metrics.

1. Download the UNBC-McMaster Shoulder Pain Dataset(http://www.consortium.ri.cmu.edu/painagree/). Delete the black files in 095-tv095\tv095t1afaff.

2. Download the pre-trained face verification model from Google Drive or Baidu Yun.

3. Detect and align the faces in the dataset by ./scripts/general_align.m.

4. Create list for Caffe's ImageData layer by ./scripts/create_list.m and ./scripts/create_sublist.m(For cross-validation).

5. Copy all the folders created by ./scripts/create_sublist.m to ./prototxt/ and run ./prototxt/run_script.cmd.

It takes about 4-6 hours to train all 25-fold cross validation.

1. Use ./scripts/extract_feature.m to extract results from the 25-fold cross-validation (because I didn't write the logic when detector cannot find face in the image).

2. Get the performance by ./scripts/get_accuracy.m.

We encourage future works to use the new proposed evaluation metrics, wMAE and wMSE. The evaluation codes are in ./scripts/get_accuracy.m.

The model definition files used in EmotionNet Challenge are provided in ./prototxt/EmotionNet. Besides the details described in our submitted paper, we further used a novel technique called mirror face to encode the features both from frontal face and mirror face, in an end-to-end fashion. The details and analysis will be described in a comming journal paper.

To try the mirror face technique, you need to compile my Caffe (https://github.com/happynear/caffe-windows/tree/ms), or transplant the flip_layer to your own Caffe.

This code is distributed under MIT LICENSE

If you find these codes useful in your research, please cite

@inproceedings{wang2017regularizing,
  title={Regularizing Face Verification Nets For Pain Intensity Regression},
  author={Wang, Feng and Xiang, Xiang and Liu, Chang and Tran, Trac D and Reiter, Austin and Hager, Gregory D and Quon, Harry and Cheng, Jian and Yuille, Alan L},
  booktitle={IEEE International Conference on Image Processing 2017},
  year={2017},
  organization={IEEE Xplore}
}

Feng Wang [feng.wff(at)gmail.com] or Xiang Xiang [xxiang(at)cs.jhu.edu]

(please replace (at) with @.)