Causal Intervention for Subject-Deconfounded Facial Action Unit Recognition

Implementation of "Causal Intervention for Subject-Deconfounded Facial Action Unit Recognition" (AAAI 2022)paper.

Dependencies

$ git clone https://github.com/echoanran/CIS.git $INSTALL_DIR

python >= 3.6
torch >= 1.1.0
requirements.txt

$ pip install -r requirements.txt

torchlight

$ cd $INSTALL_DIR/torchlight
$ python setup.py install

Data Preparation

Step 1: Download datasets

First, request for the access of the two AU benchmark datasets:

BP4D
DISFA

Step 2: Preprocess raw data

Preprocess the downloaded datasets using Dlib

Detect face and facial landmarks
Align the cropped faces according to the computed coordinates of eye centers
Resize faces to (256, 256)

Step 3: Split dataset for subject-exclusive 3-fold cross-validation

Split the subject IDs into 3 folds randomly

Step 4: Generate feeder input files

Our dataloader $INSTALL_DIR/feeder/feeder_image_causal.py requires two data files (an example is given in $INSTALL_DIR/data/bp4d_example):

label_path: the path to file which contains labels ('.pkl' data), [N, 1, num_class]
image_path: the path to file which contains image paths ('.pkl' data), [N, 1]

Training

$ cd $INSTALL_DIR
$ python run-cisnet.py

Citation

Please cite our paper if you use the codes:

@inproceedings{yingjie2022,
  title={Causal Intervention for Subject-Deconfounded Facial Action Unit Recognition},
  author={Chen, Yingjie and Chen, Diqi and Wang, Tao and Wang, Yizhou and Liang, Yun},
  booktitle={AAAI},
  year={2022}
}