A wavelet attention fusion and YOLO based approach for robust multi scale hand pose estimation

3D Hand pose estimation in multi-view 3D datasets remains a fundamental challenge in computer vision due to the complexity of human skeletal articulation, partial occlusions, variations in hand appearance and differences in human physique. Traditional models often suffer from loss of fine-grained details and poor generalization under real-world conditions. To address these limitations, we propose a novel framework for real-time 3D hand pose estimation, which integrates a Wavelet Attention Fusion Module (WAFM) with modified YOLOv11. Here, WAFM leverages wavelet decomposition to preserve high-frequency edge details while enhancing low-frequency semantic features, enabling accurate reconstruction of fine joint articulations. Our architecture extends the YOLOv11 framework by embedding its detection outputs into a multi-scale tempo-spatial backbone composed of an Efficient Net spatial extractor and Gated Recurrent Units to maintain temporal consistency across sequential frames. A specialized feature fusion neck, combining Feature Pyramid Networks and Path Aggregation Networks, ensures that high-level semantics are successfully integrated with low-level spatial details at multiple resolutions. Finally, multi-head processes these features to produce scale-aware pose estimations with improved accuracy under occlusion and varying orientations. Experimental results demonstrate that our approach significantly improves accuracy under heavy occlusion and varying orientations, achieving state-of-the-art performance in both precision and inference speed for real-time 3D hand pose reconstruction of Ego Hand dataset (Dataset I) and Senz3D dataset (Dataset II).