wallflower95 / skeleton-based-human-action-recognition

This is partial implementations of PR 2017 and CVPR 2018

3D Skeleton-based Human Action Recognition

This is partial implementations of our PR 2017 and CVPR 2018 papers.

Prerequisites

The following code is based on Matlab R2015b, Python 2.7.14, and Pytorch 0.3.0.

Installation

1. $ git clone https://github.com/nkliuyifang/Skeleton-based-Human-Action-Recognition.git

2. Download and unzip datasets:

   • UTD-MHAD
   • Northwestern-UCLA
   • NTU RGB+D

3. Open Matlab, and run "run.m"

4. $ python run.py

5. $ python show.py

Results

We report average recognition accuracy over ten times of running:

Method	UTD-MHAD Cross Subject (%)	Northwestern-UCLA Cross View (%)	NTU RGB+D Cross View (%)
Single CNN (PR 2017)	87.63	73.98	83.42
Single CNN + View Transform (PR 2017)	89.74	84.30	87.13
Pose Evolution Image (CVPR 2018)	88.84	75.65	84.72
Pose Evolution Image + View Transform	88.14	86.61	86.38

Citation

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

@article{PR 2017
    title={Enhanced Skeleton Visualization for View Invariant Human Action Recognition},
    author={Liu, Mengyuan and Liu, Hong and Chen, Chen},
    journal={Pattern Recognition (PR)},
    volume={68},
    pages={346--362},
    year={2017}
}

@inproceedings{CVPR 2018,
    title={Recognizing Human Actions as the Evolution of Pose Estimation Maps},
    author={Liu, Mengyuan and Yuan, Junsong},
    booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    pages={1159--1168},
    year={2018}
}