PCRP

Introduction

This is the official implementation of "Prototypical Contrast and Reverse Prediction: Unsupervised Skeleton based Action Recognition".

Requirements

• Python 3.6
• Pytorch 1.0.1

Datasets

• N-UCLA:
  Download transformed data from https://github.com/shlizee/Predict-Cluster/tree/master/ucla_github_pytorch/UCLAdata
• NTU RGB+D 60:
  Download transformed data from https://github.com/shlizee/Predict-Cluster.
• NTU RGB+D 120:
  Download raw data from https://github.com/shahroudy/NTURGB-D.
  Use ntu_gendata_for_predictCluster_right.py to reprocess raw data for view invariant transformation.

Put the data into the folder that matches the codes in pc_test.py

Usage

• pretrain and then linear evaluation:
  python pc_test.py

Learned Models

https://drive.google.com/drive/folders/1cck_0od9LqMIt2fvCOcca956dvCJTxcR?usp=sharing

License

PCRP is released under the MIT License.

Citation

@misc{xu2020prototypical,
title={Prototypical Contrast and Reverse Prediction: Unsupervised Skeleton Based Action Recognition},
author={Shihao Xu and Haocong Rao and Xiping Hu and Bin Hu},
year={2020},
eprint={2011.07236},
archivePrefix={arXiv},
primaryClass={cs.CV}
}