Group Graph Convolutional Networks for 3D Human Pose Estimation
Zijian Zhang (Beijing University of Posts and Telecommunications)*
The 33rd British Machine Vision Conference
Abstract
In skeleton-based 3D human pose estimation (HPE), graph convolutional networks (GCNs) have recently achieved encouraging performance. However, most previous GCNs are limited by coupling aggregation mechanism. To address this limitation, we introduce the decoupling aggregation mechanism in CNNs to GCNs and propose group graph convolutional networks (GroupGCN). It consists of two main components: group convolution and group interaction. Group convolution ensures that every group has its own spatial aggregation kernel: the adjacent matrix. Group interaction ensures that the features interact between groups. We consider four different forms of group interaction and four different types of spatial aggregation kernels, aiming to conduct a comprehensive and systematic study of decoupling aggregation mechanism in GCNs. The proposed approach achieves the state-of-the-art performance while using 70% fewer parameters.Video
Citation
@inproceedings{Zhang_2022_BMVC,
author = {Zijian Zhang},
title = {Group Graph Convolutional Networks for 3D Human Pose Estimation},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/1019.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by