Dual chain dynamic hypergraph convolution network for 3D human pose estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Graph Neural Networks

Background:

• Graph Convolutional Networks (GCNs) represent human skeletons using dynamic graphs for flexible feature aggregation.
• Optimal graph structures vary with human poses, challenging single-structure GCN approaches.
• Existing GCNs focusing on joint error loss exhibit limited adaptability and generalization.

Purpose of the Study:

• To propose a novel Dual Chain Dynamic Hypergraph Convolution Network (DCD-HCN) to address limitations in human pose estimation.
• To enhance model adaptability and generalization performance in GCNs for skeleton representation.
• To develop a method that overcomes the impracticality of estimating a single optimal graph structure for all poses.

Main Methods:

• Introduced a dual-chain structure decoupling dynamic hypergraph construction and convolution.
• Proposed an edge-weight matching mechanism to decompose hypergraph independence efficiently.
• Integrated these innovations into a Selector-Processor (SP-block) trained with supervised and unsupervised losses.

Main Results:

• The DCD-HCN achieved state-of-the-art (SOTA) generalization performance on Human3.6M and MPI-INF-3DHP datasets.
• The proposed method demonstrated competitive testing results.
• The dual-chain structure and edge-weight matching improved model adaptability.

Conclusions:

• The DCD-HCN effectively addresses the limitations of existing GCNs in human pose estimation.
• The novel framework enhances generalization by dynamically adapting graph structures.
• This research advances the field of skeleton-based human pose analysis with improved performance and adaptability.