1suancaiyu / TE-GCN

skeleton-based action recognition

TE-GCN

Code for the paper "Temporal-Enhanced Graph Convolution Network for Skeleton-based Action Recognition"

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

@article{xie_tegcn_2022,
author = {Xie, Yulai and Zhang, Yang and Ren, Fang},
doi = {https://doi.org/10.1049/cvi2.12086},
journal = {IET Comput.Vis.},
title = {{Temporal-enhanced graph convolution network for skeleton-based action recognition}},
year = {2022}
}

Note that:

• This code is based on 2s-AGCN

Data preparation

Prepare the data according to UAVHuman-Pose processing

Your data/ should be like this:

uav
___ xsub1
    ___ test_data.npy
    ___ test_label.pkl
    ___ train_data.npy
    ___ train_label.pkl
___ xsub2
    ___ test_data.npy
    ___ test_label.pkl
    ___ train_data.npy
    ___ train_label.pkl

TRAIN

You can train the your model using the scripts:

sh scripts/TRAIN_V1.sh
sh scripts/TRAIN_V2.sh

TEST

You can test the your model using the scripts:

sh scripts/EVAL_V1.sh
sh scripts/EVAL_V2.sh

WEIGHTS

We have released two trained weights in baidupan,passwd is nwhu

Your should put them into runs/.

• V1:TOP1-42.37%
• V2:TOP1-68.11%