This code provides a PyTorch implementation and pretrained models for STIGCN as described in the paper Spatio-Temporal Inception Graph Convolutional Networks for Skeleton-Based Action Recognition.
STIGCN is an efficient and simple method for skeleton-based action recognition. It overcomes the limitations of previous methods in extracting and synthesizing information of different scales and transformations from different paths at different levels (simiar to GoogLeNet). Extensive experiments demonstrate that our network outperforms state-of-the-art methods by a significant margin with only 1/5 of the parameters and 1/10 of the FLOPs.
For data preparation, please refer to 2s-AGCN for more details.
We release our model on NTU-XSub benchmark.
Change the config file depending on what you want.
`python main.py --config ./config/ntu-xsub.yaml`
`python main.py --config ./config/test.yaml`
Please cite the following paper if you use this repository in your reseach.
@inproceedings{huang2020spatio,
title={Spatio-Temporal Inception Graph Convolutional Networks for Skeleton-Based Action Recognition},
author={Huang, Zhen and Shen, Xu and Tian, Xinmei and Li, Houqiang and Huang, Jianqiang and Hua, Xian-Sheng},
booktitle={ACM MM},
pages={2122--2130},
year={2020}
}