dangxusheng/jaanet

ECCV 2018 "Deep Adaptive Attention for Joint Facial Action Unit Detection and Face Alignment"

JAANet

This repository implements the training and testing of JAA-Net for "Deep Adaptive Attention for Joint Facial Action Unit Detection and Face Alignment". It offers the original implementation of the paper in Caffe. We also release the PyTorch implementation here

Getting Started

Dependencies

• Dependencies for Caffe are required

• The new implementations in the folders "src" and "include" should be merged into the official Caffe:

  • Add the .cpp, .cu files into "src/caffe/layers"
  • Add the .hpp files into "include/caffe/layers"
  • Add the content of "caffe.proto" into "src/caffe/proto"
  • Add "tools/convert_data.cpp" into "tools"

• New implementations used in our paper:

  • au_mask_based_land_layer: generate attention maps given the locations of landmarks
  • division_layer: divide a feature map into multiple identical subparts
  • combination_layer: combine mutiple sub feature maps
  • data_layer and data_transform_layer: the processing of landmarks in the case of mirroring faces is added
  • align_data_transform_layer: reset the order and change the coordinates for landmarks in the cases of mirroring and cropping
  • dice_coef_loss_layer: Dice coefficient loss
  • softmax_loss_layer: the weighting for the loss of each element is added
  • euclidean_loss_layer: the weighting for the loss of each element and the normalizing with inter-ocular distance are added
  • convert_data: convert the AU and landmark labels, weights, and reflect_49 to leveldb or lmdb

• Build Caffe

Datasets

BP4D and DISFA

The 3-fold partitions of both BP4D and DISFA are provided in the folder "data", in which the path file of DISFA contains each frame of videos

Preprocessing

• Prepare the training data
  • Run "prep/face_transform.cpp" to conduct similarity transformation for face images
  • Run "prep/write_biocular.m" to compute the inter-ocular distance of each face image
  • Run "prep/combine2parts.m" to combine two partitions as a training set, respectively
  • Run "prep/write_AU_weight.m" to compute the weight of each AU for the training set
  • Run "tools/convert_imageset" of Caffe to convert the images to leveldb or lmdb
  • Run "tools/convert_data" to convert the AU and landmark labels, inter-ocular distances, weights and reflect_49 to leveldb or lmdb: the example format of files for AU and landmark labels, inter-ocular distances and weights are in "data/examples"; reflect_49.txt is in "data"; the weights are shared by all the training samples (only one line needed); reflect_49 is used to reset the order and change the coordinates for landmarks in the case of face mirroring
  • Our method is evaluated by 3-fold cross validation. For example, “BP4D_combine_1_2” denotes the combination of partition 1 and partition 2
• Modify the "model/BP4D_train_val.prototxt":
  • Modify the paths of data
  • A recommended training strategy is that selecting a small set of training data for validation to choose a proper maximum iterations and then using all the training data to retrain the model
  • The loss_weight for DiceCoefLoss of each AU is the normalized weight computed from the training data
  • The lr_mult for "au*_mask_conv3*" corresponds to the enhancement coefficient "\lambda_3", and the loss_weight of "au*_mask_loss" is related to the reconstruction constraint "E_r" and "\lambda_3"

   When \lambda_3 = 1:
   
    param {
      lr_mult: 1
      decay_mult: 1
    }
    param {
      lr_mult: 2
      decay_mult: 0
    }
    
    loss_weight: 1e-7

   When \lambda_3 = 2:
   
    param {
      lr_mult: 2
      decay_mult: 1
    }
    param {
      lr_mult: 4
      decay_mult: 0
    }
    
    loss_weight: 5e-8

• There are two minor differences from the original paper without performance degradation:
  • The redundant cropping of attention maps is removed
  • The first convolution of the third block uses the stride of 2 instead of 1 for smaller model complexity

Training

cd model
sh train_net.sh

• Trained models on BP4D with 3-fold cross-validation can be downloaded here

Testing

• Compute evaluation metrics

python test.py

• Visualize attention maps

python visualize_attention_map.py

Citation

If you use this code for your research, please cite our paper

@inproceedings{shao2018deep,
  title={Deep Adaptive Attention for Joint Facial Action Unit Detection and Face Alignment},
  author={Shao, Zhiwen and Liu, Zhilei and Cai, Jianfei and Ma, Lizhuang},
  booktitle={European Conference on Computer Vision},
  year={2018},
  pages={725--740},
  organization={Springer}
}

Acknowledgments

Code is partially inspired by DRML and A-Variation-of-Dice-coefficient-Loss-Caffe-Layer