HandAttNet: Attention 3D Hand Mesh Estimation Network

Jintao Sun, Gangyi Ding

Abstract: Hand pose estimation and reconstruction become increasingly compelling in the metaverse era. But in reality hands are often heavily occluded, which makes the estimation of occluded 3D hand meshes challenging.Previous work tends to ignore the information of the occluded regions, we believe that the occluded regions hand information can be highly utilized, Therefore, in this study, we propose hand mesh estimation network, HandAttNet.We design the cross-attention mechanism module and the DUO-FIT module to inject hand information into the occluded region.Finally, we use the self-attention regression module for 3D hand mesh estimation.Our HandAttNet achieves SOTA performance.