3D Human Pose Estimation via Graph Extended Spatio-Temporal Convolutional Network

Abstract

3D human pose estimation is an important premise for human behavior analysis and understanding, which has a wide range of applications in intelligent transportation, human-computer interaction, and animation production. Most existing works focus on extracting the feature relationship between frames by combining spatio-temporal information to reduce the error of attitude reconstruction. However, the majority of them often suffer from insufficient joint correlation characteristics. To address this problem, we propose a Graph Expand Spatiotemporal Convolutional Network, named GESC-Net, to improve the limitation of extracting human spatial structure features. To better enrich the feature of extracting local information, we develop a learnable symmetric connection (LSC) block in the spatial structure. Moreover, a CbAttantion block is also designed to obtain a larger view of the acquisition of global structure and get more effective features. We evaluate our approach on two standard benchmark datasets: Human3.6M and HumanEva-I. The quantitative and qualitative evaluation results demonstrate that the GESC-Net can achieve better 3D human posture estimation than existing state-of-the-art methods.