We provide the training and testing codes for SAN, implemented in PyTorch.

• Python3.6
• PyTorch
• torchvision

• Download 300W-Style and AFLW-Style from Google Drive, and extract the downloaded files into ~/datasets/.
• In 300W-Style and AFLW-Style directories, the Original sub-directory contains the original images from 300-W and AFLW

cd cache_data
python aflw_from_mat.py
python generate_300W.py

The generated list file will be saved into ./cache_data/lists/300W and ./cache_data/lists/AFLW.

python crop_pic.py

The above commands are used to pre-crop the face images.

• Step-1 : cluster images into different groups, for example sh scripts/300W/300W_Cluster.sh 0,1 GTB 3.
• Step-2 : use sh scripts/300W/300W_CYCLE_128.sh 0,1 GTB or sh scripts/300W/300W_CYCLE_128.sh 0,1 DET to train SAN on 300-W.

• Step-1 : cluster images into different groups, for example sh scripts/AFLW/AFLW_Cluster.sh 0,1 GTB 3.
• Step-2 : use sh scripts/AFLW/AFLW_CYCLE_128.FULL.sh or sh scripts/AFLW/AFLW_CYCLE_128.FRONT.sh to train SAN on AFLW.

Please cite the following paper in your publications if it helps your research:

@inproceedings{dong2018san,
   title={Style Aggregated Network for Facial Landmark Detection},
   author={Dong, Xuanyi and Yan, Yan and Ouyang, Wanli and Yi, Yang},
   booktitle={Computer Vision and Pattern Recognition},
   year={2018},
}

To ask questions or report issues, please open an issue on the issues tracker.