GMDN: A lightweight graph-based mixture density network for 3D human pose regression. (April 2021)

• Record Type: Journal Article
• Authors: Zou, Lu; Huang, Zhangjin; Gu, Naijie; Wang, Fangjun; Yang, Zhouwang; Wang, Guoping
• Is Part Of: Computers & graphics. Volume 95(2021)
• Page Start: 115
• Page End: 122
• Publication Date: 2021-04
• Subjects: 3D human pose estimation -- Graph convolutional network -- Mixture density network; Computer graphics -- Periodicals; 006.6
• DOI: 10.1016/j.cag.2021.01.010
• Languages: English
• ISSNs: 0097-8493

Abstract

Highlights: We formulate the 2D joint locations of the human body as a graph and present a novel lightweight graph convolutional operation with structural knowledge about human bodies. Based on the proposed graph convolutional operation, a novel graph-based mixture density network (GMDN) is proposed to resolve the ambiguity and occlusion problem existing in 3D human pose estimation. Comprehensive experiments on the Human3.6M dataset demonstrate that GMDN achieves state-of-the-art performance with only 0.30M parameters. Graphical abstract: Abstract: 3D human pose estimation from 2D detections is an ill-posed problem because multiple solutions may exist due to the inherent ambiguity and occlusion. In this paper, we propose a novel graph-based mixture density network (GMDN) to tackle the 2D-to-3D human pose estimation problem. We formulate the 2D joint locations of the human body as a graph, and thus the pose estimation task can be redefined as a graph regression problem. Additionally, we present a novel graph convolutional operation with the incorporation of structural knowledge about human body configurations to assist with reasoning of the structural relations implied in the human bodies. Furthermore, we employ mixture density networks to formulate the 3D human poses as a multimodal distribution. The presented GMDN is lightweight with only 0.30M parameters, and the experimental results demonstrate that it achieves state-of-the-art performance.