Part-based Graph Convolutional Network for Skeleton-based Action Recognition

Official repository for the code from BMVC (British Machine Vision Conference) paper "Part-based Graph Convolutional Network for Action Recognition". The implementation is done in Pytorch and works on it's recent stable version. The repository includes:

• Code for the final model used in the paper.
• Model checkpoints for model trained on NTURGB+D Cross Subject and Cross View data splits.
• Training and testing config as well as data preparation scripts for NTURGB+D dataset.
• Training config as well as data preparation scripts for HDM05 dataset.

TODOs:

• Code for visualizing results.
• Document how to extend the code for other projects.

Getting Started

1. Download the NTURGB+D dataset (with 60 action classes) following this link. Unzip the archive and store all the skeleton files in a single directory:

unzip nturgbd_skeletons_s001_to_s017.zip -d nturgb+d_skeletons

2. Clone the repository:

git clone https://github.com/kalpitthakkar/pb-gcn.git

3. Download the pretrained model checkpoints. To download the checkpoints for both the cross-subject and cross-view splits:

bash download_checkpoints.sh <path_to_download_directory>

Training / Testing Models

1. First of all, we need to define the required configuration variables in the YAML file. The instructions on it's structure and editing the file are here.

2. Once the configuration file is ready, you can start training the model:

python run.py --config <path_to_YAML_config_file>

Citation

For citing our paper:

@article{thakkar2018part,
title={Part-based Graph Convolutional Network for Action Recognition},
author={Thakkar, Kalpit and Narayanan, PJ},
journal={arXiv preprint arXiv:1809.04983},
year={2018}
}