DirecFormer: A Directed Attention in Transformer Approach to Robust Action Recognition

This is the official implementation of the paper "DirecFormer: A Directed Attention in Transformer Approach to Robust Action Recognition".

Installation

The installation of this repository is similar to the installation of SlowFast. The instruction can be found here

To prepare a dataset, you should follow the instructions here provided by SlowFast.

Testing

To test the model on the Jester dataset, you can perform the following commands:

python tools/run_net_tsm.py --cfg config/Jester/DirecFormer.yaml \
            TRAIN.ENABLE False \
            TEST.CHECKPOINT_FILE_PATH <PATH-TO-CHECKPOINT> \

Training and Optimization

Please contact the Project Investigator (Khoa Luu) for further information about training models, optimized models on-the-edge and low-cost devices.

Acknowledgements

This codebase is borrowed from SlowFast and TimeSFormer

Citation

If you find this code useful for your research, please consider citing:

@inproceedings{truong2021direcformer,
  title={DirecFormer: A Directed Attention in Transformer Approach to Robust Action Recognition},
  author={Truong, Thanh-Dat and Bui, Quoc-Huy and Duong, Chi Nhan and Seo, Han-Seok and Phung, Son Lam and Li, Xin and Luu, Khoa},
  booktitle={Computer Vision and Pattern Recognition},
  year={2022}
}