Abstract: There are enormous challenges for 3D pose estimation due to non-rigid characteristics and occlusion. And modeling joint relationship is a key technology for that. It is difficult to deal with joint missing or skewing caused by occlusion or singularity if only modeling static joint relationships from human physical construction. It is beneficial to improve 3D pose estimation when extracting pose semantics and modeling dynamic joint relationship but the wholeness and hierarchy on human joint relationship can’t be neglected. This work proposes modeling static and dynamic joint relationships together for 3D pose estimation. Mutual information algorithm is used for obtaining a joint relationship map which used to group human joints. Then accumulate the human joint groups based on three level human joint freedoms level by level. A cascading estimation and group joint feature sharing networks are designed to model static joint relationship. Multi-group attention mechanisms are present for each level for extrating pose semantic feature to model dynamic joint relationships. A data enhancement policy by category balancing and pose reorganizing is presented further for improving model robustness. The experiments are carried out based on Human3.6M, MPI-INF-3DHP and MPII datasets extensively. The results show that the average error of our model is reduced 0.2 mm and the average accuracy is increased 0.2% at least when comparing with other advanced models. And the model performance is improved significantly when our data enhancement policy is carried out.

Key words: depth ambiguity, joint relationship modeling, attention mechanism, 3D human pose estimation