A Global-Part-Local Approach for 3D Human Pose Estimation from Single-View Images
Yuhong Xie, Chaoqun Hong, Rongsheng Xie, Jie Li · 2023
Significant progress has been made in 3D human pose estimation (HPE) from monocular images. Previous research has shown that in the process of 3D HPE, global information provides an overall spatial structure and a rough layout of the pose. In contrast, local information plays a crucial role in precisely locating specific body parts. The combined consideration of both aspects proves advantageous for effectively acquiring a comprehensive representation of the human skeleton. However, the human body possesses an intrinsic topological structure, which gives rise to pose estimation errors. These errors have a tendency to propagate within the body's interconnected framework, causing them to accumulate predominantly at the distal joints. This accumulation, over time, culminates in a discernible reduction in the overall accuracy of pose estimation. To address this issue, we propose a method termed GlPaLo (Global-Part-Local), which integrates global, part-level, and local information. It consists of two key modules: uMLPGraph algorithm and BPConstraint module. GlPaLo aims to capture global, part, and local information among human keypoints to improve the accuracy of 3D HPE. Our uMLPGraph algorithm module consists of a multi-layer perceptron with a U-shaped structure (uMLP) and a graph convolutional network (GCN), which is used to simultaneously process both global and local information. The BPConstraint module is divided into body-level constraints and part-level constraints, aiming to learn constraint information about the human body structure. At the part-level constraints, introducing parent node features as prior knowledge helps to reduce the accumulation of errors at the end joints of the human body. Extensive experiments conducted on the Human3.6M and MPI-INF-3DHP datasets demonstrate that the proposed method achieves outstanding performance.