Multi-view fusion with adaptive feature integration for skeleton-based action recognition
Year: 2026
Type: article
Abstract: GCN-based skeleton action recognition methods have long relied on physical connections between joints to construct initial topological structures. However, human actions are determined not only by pairwise bone connections but also by coordinated movements among multiple joints across different body parts. This makes a single binary topology insufficient to capture diverse inter-joint dependencies, especially higher-order relations beyond physical connections. To alleviate the above bottlenecks, we propose the multi-view fusion network (MVF-Net). First, MVF-Net introduces an MVF paradigm to construct a unified topology from complementary structural views, including hypergraph-induced higher-order relations, physical skeleton connections, and local neighborhood structures. The resulting topology serves as a structured spatial prior for graph convolution and is further combined with learnable relation modeling, enabling more comprehensive and expressive feature propagation across joints. To better exploit the resulting spatial features in the temporal domain, we further introduce a main-residual adaptive fusion (MR-AF) module that jointly captures local detail cues and global contextual semantics. MR-AF replaces fixed residual addition with a fine-grained AF strategy, dynamically regulating the relative contributions of main-path temporal features and residual information. Experiments on three large-scale benchmark datasets demonstrate that MVF-Net achieves competitive and stable performance compared with recent skeleton-based action recognition methods, while offering a novel perspective for modeling data with complex topological relationships.
Source: Engineering Research Express
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,400
Award: 基于图小波的几何深度学习及其在3D点云中的应用研究
OpenAlex ID: W7164903085