PR-GCN

Introduction

This repository provides the official PyTorch implementation for the following paper:

Pose Refinement Graph Convolutional Network for Skeleton-basedAction Recognition
Shijie Li, Jinhui Yi, Yazan Abu Farha, Juergen Gall
In RA-L with ICRA 2021 .
Paper

Abstract: *With the advances in capturing 2D or 3D skeleton data, skeleton-based action recognition has received an increasing interest over the last years. As skeleton data is commonly represented by graphs, graph convolutional networks have been proposed for this task. While current graph convolutional networks accurately recognize actions, they are too expensive for robotics applications where limited computational resources are available. In this paper, we therefore propose a highly efficient graph convolutional network that addresses the limitations of previous works. This is achieved by a parallel structure that gradually fuses motion and spatial information and by reducing the temporal resolution as early as possible. Furthermore, we explicitly address the issue that human poses can contain errors. To this end, the network first refines the poses before they are further processed to recognize the action. We therefore call the network Pose Refinement Graph Convolutional Network. Compared to other graph convolutional networks, our network requires 86%-93% less parameters and reduces the floating point operations by 89%-96% while achieving a comparable accuracy. It therefore provides a much better trade-off between accuracy, memory footprint and processing time, which makes it suitable for robotics applications. *

Instalation:

python setup.py develop

Basic usage:

Any application in mmskeleton is described by a configuration file. That can be started by a uniform command:

python run.py $CONFIG_FILE [--options $OPTHION]

which is equivalent to

mmskl $CONFIG_FILE [--options $OPTHION]

Optional arguments options are defined in configuration files. You can check them via:

mmskl $CONFIG_FILE -h

Data Preparation

We experimented on two skeleton-based action recognition datasts: Kinetics-skeleton and NTU RGB+D. Before training and testing, for the convenience of fast data loading, the datasets should be converted to the proper format. Please download the pre-processed data from GoogleDrive and extract files with

cd pr-gcn
unzip <path to pr-gcn-processed-data.zip>

Training

To train a PR-GCN model, run

mmskl configs/recognition/pr_gcn/$DATASET/train.yaml [optional arguments]

The usage of optional arguments can be checked via adding --help argument. All outputs (log files and ) will be saved to the default working directory. That can be changed by modifying the configuration file or adding a optional argument --work_dir $WORKING_DIRECTORY in the command line.

Evaluation

After that, evaluate your models by:

mmskl configs/recognition/pr_gcn/$DATASET/test.yaml --checkpoint $CHECKPOINT_FILE

License

The project is release under the Apache 2.0 license.

Acknowledgments

The code is greatly inspired by mmskeleton

Citation

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

@ARTICLE{9345415,
  author={S. {Li} and J. {Yi} and Y. {Abu Farha} and J. {Gall}},
  journal={IEEE Robotics and Automation Letters}, 
  title={Pose Refinement Graph Convolutional Network for Skeleton-based Action Recognition}, 
  year={2021},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/LRA.2021.3056361}}

Contact

For any question, feel free to contact

Shijie Li     : lishijie@iai.uni-bonn.de