Abstract
Accurate 3D hand reconstruction remains a challenging computer vision problem with numerous applications. Existing approaches predominantly focus on single-frame hand reconstruction, thereby neglecting crucial temporal information essential for stable hand tracking. A novel pipeline for 3D hand reconstruction from consecutive multi-view depth images, termed HandFlowNet, is presented in this work. The proposed methodology converts multi-view depth images into a single point cloud, from which temporal information between consecutive frames is deduced through estimated scene flow of hand mesh vertices. The scene flow estimator establishes one-to-one correspondences between point sets from sequential depth frames. This scene flow is subsequently applied as an offset to initially estimated hand mesh vertices from the previous frame to determine the current frame's hand mesh vertices. These vertices are further refined using a graph convolutional network that incorporates predicted local and global features of the current frame. Through extensive evaluations, HandFlowNet is demonstrated to achieve state-of-the-art performance on public real hand benchmarks including DexYCB and HO3D.