Multi-Scale Large Kernel Attention Network for Single-View 3D Hand Mesh Reconstruction

Single-view 3D hand mesh reconstruction from RGB images is crucial for human-computer interaction. However, severe self-occlusions and hand-object occlusions frequently cause erroneous pose estimation and mesh distortions in reconstructed results. Existing CNN-based methods with fixed receptive fields struggle to adapt to diverse occlusion scenarios, while Transformer-based approaches suffer from invalid similarity computation in self-attention mechanisms due to feature ambiguity within occluded regions, consequently failing to effectively leverage object shape and positional cues for supplementing obscured hand features. To address these challenges, we propose HandMAN, a Multi-Scale Large-Kernel Attention Network for modeling hand-object interactions. Our architecture incorporates three parallel large kernel attention branches that adaptively extract both local hand details and global contextual hand-object interaction features through multi-scale receptive fields. This design preserves fine-grained local information while enabling efficient long-range interaction modeling. Furthermore, we integrate spatial and channel attention mechanisms to refine and enhance aggregated features. Extensive experiments demonstrate HandMAN's superior performance in 3D hand mesh reconstruction under challenging occlusion scenarios.