Abstract
Recently, various Graph Convolution Networks (GCNs) have been developed to represent the human skeleton using dynamic graph structures, enhancing the flexibility of node feature aggregation. However, the optimal graph structure can vary significantly across different human poses, making it impractical to estimate a single optimal graph structure for all poses. Previous experiments have also shown that the optimal graph structures derived by existing GCN methods, which focus solely on joint error loss, tend to reduce model adaptability. This limitation further compromises the generalization performance of the models. To address this issue, we propose a novel Dual Chain Dynamic Hypergraph Convolution Network (DCD-HCN). This framework introduces a dual-chain structure that decouples the processes of dynamic hypergraph construction and hypergraph convolution. Additionally, we propose a new edge-weight matching mechanism to decompose the independence of hypergraphs into the independence of hyperedges with low computational complexity. These two innovations are integrated into a Selector-Processor block (SP-block) within the DCD-HCN, which is trained with both supervised joint error loss and unsupervised extra hypergraph construction loss. Experiments on the Human3.6M and MPI-INF-3DHP datasets demonstrate that our method achieves state-of-the-art (SOTA) generalization performance while maintaining competitive testing results.