Enhanced Limb Constraints and Global Attention Integration for 3D Human Pose Estimation
Binzi Wang, Yang Shu, Xiangrong Fu, Youdong Ding · 2024
Handling single-frame 2D-3D lifting methods entails significant uncertainty. To mitigate this challenge, we propose a novel cross-parallel model named Spatial Collaboration Network (SCNet), which fully exploits the natural structural dependencies of the human skeletal system. Specifically, we amalgamate the benefits of Graph Convolutional Networks (GCN) and Transformers by formulating the Limb Constraints (LCT) module grounded in GCN and the Global Attention (GAT) module based on Transformer. Considering the joint coordination in the human body, we introduce additional connections for joints with equal degrees of freedom. Simultaneously, to alleviate the accumulation of errors at the extremities of limbs, we design an Enhanced Limb Constraints (EnhancedLCT) module. Through this progressive constraint strategy for local joints, the model better learns limb motion characteristics. Considering the interaction between local and global information, we design a Spatial Collaboration Module (SCM) to enable the model to adequately capture multi-level dependency relationships. Comprehensive experimentation reveals the outstanding performance of SCNet on two demanding datasets (Human3.6M and MPI-INF-3DHP).