MAC-Learning

Multi-Granularity Anchor-Contrastive Representation Learning for Semi-supervised Skeleton-based Action Recognition

Requirements

python == 3.8.3
pytorch == 1.11.0
CUDA == 11.2

Data Preparation

Download the raw data of NTU-RGB+D, NW-UCLA, and Skeleton-Kinetics. Then the commands for data preprocessing are as follows,

python ./data_gen/ntu_gendata.py
python ./data_gen/ucla_gendata.py
python ./data_gen/kinetics_gendata.py

Training

On NTU RGB+D cross-subject benchmark.

python main.py --config ./config/nturgbd-cross-subject/train_joint_aagcn.yaml

On NTU RGB+D cross-view benchmark.

python main.py --config ./config/nturgbd-cross-view/train_joint_aagcn.yaml

On NW-UCLA.

python main.py --config ./config/ucla/train_joint_aagcn.yaml

On Skeleton-Kinetics.

python main.py --config ./config/kinetics-skeleton/train_joint_aagcn.yaml

Trained model

The trained weight is here

The corresponding processed data (data_CS5) is here

python main.py --config ./config/nturgbd-cross-subject/train_joint_aagcn.yaml --weights ./runs/ntu_cs_aagcn_joint_best.pt --phase test

Acknowledgements

This repo is based on 2s-AGCN, thanks to the original authors for their works!

Citation

Please cite the following paper if you use this repository in your reseach.

@article{shu2022multi,
  title={Multi-Granularity Anchor-Contrastive Representation Learning for Semi-Supervised Skeleton-Based Action Recognition},
  author={Shu, Xiangbo and Xu, Binqian and Zhang, Liyan and Tang, Jinhui},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2022}
}