Multi-hop Modulated Graph Convolutional Networks for 3D Human Pose Estimation
Jae Yung Lee (KT (Korea Telecom)),* IGIL KIM (KT)
The 33rd British Machine Vision Conference
Abstract
Graph convolutional networks (GCNs) have recently been applied to 3D human pose estimation (3D HPE) from 2D body joints. GCN-based 3D HPE has achieved promising performance in modeling the relationships between body parts. However, vanilla graph convolution only considers the relationship between neighbouring joints at a one-hop distance. Some recent approaches have utilised high-order graph convolution to model long-range dependency. They exploit the adjacency matrix of one-hop neighbouring joints, but the method cannot capture the long-range dependency. To solve this problem, we propose the multi-hop modulated GCN (MM-GCN) for 3D HPE. The unique adjacency matrix of each hop distance is derived, and aggregate features of nodes at various hop distances are modulated to capture the long-range dependency. Thus, the proposed network can model a wide range of interactions between body joints more adequately than does the vanilla graph approach. Moreover, we investigate the impact of combination with affinity modulation (AM) because AM adjusts the graph in a GCN. Our experiments, and an ablation study conducted on two standard benchmarks demonstrate the effectiveness of the proposed network, showing that our MM-GCN outperforms some recent state-of-the-art techniques.Video
Citation
@inproceedings{Lee_2022_BMVC,
author = {Jae Yung Lee and IGIL KIM},
title = {Multi-hop Modulated 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/0207.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by