PainRegression

Regularizing Face Verification Nets For Pain Intensity Regression

https://arxiv.org/abs/1702.06925

#Requirements

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).

#Regression Results and Evaluation

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.

#Training

1. Download the UNBC-McMaster Shoulder Pain Dataset(http://www.consortium.ri.cmu.edu/painagree/).

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.

#Validation

1. Use ./scripts/extract_feature.m to extract results from the 25-fold cross-validation.

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

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

#License

This code is distributed under MIT LICENSE

#Contact

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

(please replace (at) with @.)