avijit9 / HiGCIN

Pytorch implementation of the TPAMI paper of "HiGCIN: Hierarchical Graph-based Cross Inference Network for Group Activity Recognition"

HiGCIN

Pytorch implementation of HiGCIN: Hierarchical Graph-based Cross Inference Network for Group Activity Recognition. (IEEE T-PAMI, 2020)

Get started

Prerequisite

Our approach is tested on only Ubuntu with GPU and it needs at least 16G GPU memory. The neccseearay packages can be installed by the following commonds:

conda create -n HiGCIN python=3.6
conda activate HiGCIN
pip install cmake dlib scikit-image sklearn h5py
pip install torch torchvision

Preprocess datasets

Download two datasets (i.e., Volleyball Dataset and Collective Activity Dataset) and unzip them to './dataset/VD/videos' and './dataset/CAD/videos', respectively. Then run the following commands:

bash pre_script.sh 'VD'
bash pre_script.sh 'CAD'

Alternatively, you can also direct download the personal tracklets from here (code is o8fy) and put them in './dataset/VD/imgs' and './dataset/CAD/imgs', respectively.

Train a Standard Model from Scratch

bash traintest_script.sh

Citation

If you wish to refer to the results of HiGCIN, please use the following BibTeX entry.

@article{yan2020higcin,
  title={HiGCIN: Hierarchical Graph-based Cross Inference Network for Group Activity Recognition},
  author={Yan, Rui and Xie, Lingxi and Tang, Jinhui and Shu, Xiangbo and Tian, Qi},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2020},
  doi={10.1109/TPAMI.2020.3034233}
}

Update Records

12/3/2020: Volleyball Dataset is supported now.

Todo

• Support Collective Activity Dataset

Acknowledgments

Thanks to the pytorch version implementation of Non-Local from https://github.com/AlexHex7/Non-local_pytorch

Contact Information

Feel free to create a pull request or contact me by Email = ["ruiyan", at, "njust", dot, "edu", dot, "cn"], if you find any bugs. For further information about me, welcome to my homepage.