Im2Hands: Learning Attentive Implicit Representation of Interacting Two-Hand Shapes

Image alignment Reconstruction Zoom-in Input Image alignment Reconstruction Zoom-in IntagHand Im2Hands (ours) Observation Reconstruction Image alignment Image alignment Reconstruction IntagHand [23] Im2Hands (Ours) Figure 1. Reconstructed two-hand shapes from an RGB image. We present Im2Hands, the first neural implicit representation for two interacting hands. Compared to the existing mesh-based two-hand reconstruction method (IntagHand [23]), Im2Hands effectively captures the fine-grained geometry of two hands with higher shape-to-image coherency. The above results were produced from a single RGB image input, where ours utilized an off-the-shelf two-hand keypoint estimation method (DIGIT [11]) to condition the articulated occupancy.