Global Relation Reasoning Graph Convolutional Networks for Human Pose Estimation

Abstract

We explore the importance of global relation reasoning in Human Pose Estimation (HPE). Global relation reasoning aims to globally learn relations among regions of images or videos. For HPE, if we can globally model the relations among different body joints, we may mitigate some challenges such as occlusion. Most existing human pose estimation methods rely on building Convolutional Neural Networks (CNNs). Because convolution operations can only model local relations, in order to capture global relations, they must inefficiently stack multiple convolution layers to enlarge the receptive fields to cover all the body joints in the image. In this paper, we propose to utilize Global Relation Reasoning Graph Convolutional Networks (GRR-GCN) to efficiently capture the global relations among different body joints. GRR-GCN projects all the features in the original coordinate space to a graph space. In the graph space, these features are represented by a set of nodes to form a fully-connected graph, on which global relation reasoning is performed by graph convolution. After reasoning, node features are projected back to the coordinate space for further processing. GRR-GCN is a plug-and-play module, and can be integrated into current state-of-the-art networks. Experiments on human pose estimation benchmark, MPII and COCO keypoint detection dataset, show that GRR-GCN can boost the performance of state-of-the-art human pose estimation networks including SimpleBaseline and HRNet (High-Resolution Net).

Keywords

• Human pose estimation
• graph convolutional networks (GCN)
• global relation reasoning