Hierarchical hypergraph parallel networks for 3D human pose estimation
shulin dai, Xiuning Chen, Wei Xia, Huabin Wang · 2025
Graph convolutional networks (GCNs) play a critical role in 3D human pose estimation by modeling the human skeleton structure as a graph, which allows for capturing spatial relationships between joints and learning effective representations of underlying poses. However, studies have revealed a potential limitation of existing graph-based models: pairwise relationships inherently overlook high-order kinematic dependencies between body joints, making it challenging to accurately capture the complex and varying relationships between joints. To address this issue, a hypergraph representation method based on a hierarchical hyper-neighborhood construction is proposed. This method incorporates potential joint relationships into the same hyperedge of a hypergraph, which not only considers pairwise relationships but also captures multi-level dependencies among different human motion chains.Additionally, we designed a parallel network that learns different features by partitioning channels, progressively capturing information about the human skeletal topology from local joints to the global body at multiple levels. This approach effectively aggregates joint features across different channels, thereby achieving human pose estimation based on joint information. The proposed model was evaluated on two standard benchmark datasets, Human3.6M and MPI-INF-3DHP, and compared against a series of state-of-the-art methods for 3D human pose estimation. Experimental results demonstrate that our method achieves superior performance on both datasets. Furthermore, ablation studies were conducted to analyze the contributions of different components in the model architecture, showing that hypergraph convolution significantly improves the model's performance.