Total: 1
Parse graphs have been widely used in human pose estimation (HPE) to model the hierarchical structure and context relations of the human body, which has been shown to effectively improve robustness and accuracy. But most methods rely on parse graphs built from predefined skeletons, causing two key issues: poor integratability with other models, and complex designs with redundant parameters for hierarchy and context relation modeling of body. To address these issues, we propose a novel Refinement Module based on Parse Graph (RMPG). RMPG abandons skeleton connections and refines features by building implicit hierarchical structures and context relations between sub-feature maps, with strong integratability. Furthermore, our hierarchical network design demonstrates that RMPG can model the body's hierarchical structure and context relations with a simpler architecture and fewer parameters. RMPG operates in two stages: the top-down decomposition recursively partitions the feature map into a tree-structured hierarchy, where each node corresponds to a sub-feature map; the bottom-up composition aggregates context information to progressively refine the feature representation. Extensive experiments show that RMPG can be flexibly embedded into various methods, including our hierarchical networks, and consistently improves performance across multiple mainstream HPE benchmarks. The code will be released.