问题

The paper addresses the underutilization of spatio-temporal body structure features in transformers and inadequate granularity of spatio-temporal interaction modeling in graph convolutional networks for 3D human pose estimation, leading to depth ambiguity.

方法

The study introduces the Spatio-Temporal GraphFormer (STGFormer) framework, featuring a Spatio-Temporal criss-cross Graph attention mechanism and a dual-path Modulated Hop-wise Regular GCN to process temporal and spatial dimensions independently.

关键发现

The proposed method achieves state-of-the-art performance on the Human3.6M and MPI-INF-3DHP datasets, demonstrating superior capacity for learning body structure and capturing long-range dependencies compared to transformer-based, GCN-based, and hybrid methods.

3个要点

• STGFormer effectively integrates spatio-temporal graph structures into attention layers, leveraging prior knowledge for improved long-range dependency capture.
• The dual-path modulated hop-wise regular GCN module extracts first-order and higher-order spatio-temporal graph information from features efficiently.
• Extensive experiments confirm STGFormer's efficacy and generalization, outperforming state-of-the-art techniques on benchmark datasets for 3D human pose estimation.