startime-h/drml

Deep Region and Multi-label Learning for Facial Action Unit Detection (CVPR16)

Intro

This repository provides the codes for the CVPR16 paper, “Deep Region and Multi-Label Learning for Facial Action Unit Detection". This code aims for training a convolutional network that contains a region layer for specializing the learned kernels on different facial regions, and meanwhile utilizes a multi-label cross-entropy to jointly learn 12 AUs. This implementation is based on Caffe Toolbox.

File structure

Based on the caffe toolbox, we organize the source files as follows:

• include/caffe/: Header files that contains the declaration of our implemented layers

• prototxt/: Network architecture we used to compuare and report in our paper

• src/caffe/layers/: Source files of our implemented layers

  • box_layer.*: Slice a 160x160 response map into an 8x8 uniform grid.

  • image_data_layer_multilabel.cpp: Load multiple labels for one image.

  • multi_sigmoid_cross_entropy_loss_layer.*: Multi-label loss.

  • splice.*: Concatenate 20 8x8 uniform grids to a 160x160 feature map.

More info

• Contact: Please send comments to Kaili Zhao ([email protected])
• Citation: If you use this code in your paper, please cite the following:

@inproceedings{zhao2016deep,
  title={Deep Region and Multi-Label Learning for Facial Action Unit Detection},
  author={Zhao, Kaili and Chu, Wen-Sheng and Zhang, Honggang},
  booktitle={CVPR},
  year={2016}
}