HOGFormer: high-order graph convolution transformer for 3D human pose estimation

Abstract

The combination of graph convolution network (GCN) and Transformer has shown promising results in 3D human pose estimation (HPE) tasks when lifting the 2D to 3D poses. However, recent approaches to 3D HPE still face difficulties such as depth ambiguity and occlusion. To address these issues, we suggest a novel 3D HPE architecture, termed High-Order Graph Convolution Transformer (HOGFormer). HOGFormer consists of three core components: the Chebyshev Graph Convolution (CGConv) module, the Graph-based Dynamic Adjacency Matrix Transformer (GDAMFormer) module, and the High-Order Graph Convolution (HOGConv) module. In more detail, the CGConv module can further increase the estimation accuracy by approximating the graph convolution with Chebyshev polynomials. The GDAMFormer module efficiently addresses issues like self-occlusion and depth blur by using a dynamic adjacency matrix to represent the dynamic relationships among joints. The HOGConv module can effectively extract local features by capturing the local physical dependencies of skeleton connections. With the integration of these modules, the proposed architecture can effectively capture global and local information. We evaluate our architecture quantitatively and qualitatively on the popular benchmark dataset Human3.6M. Our experiments demonstrate that HOGFormer achieves state-of-the-art performance.