Camera-Space Hand Mesh Reconstruction from a Monocular Image via Pseudo Stereo Perception

Recovering a reliable three-dimensional (3D) hand mesh from a monocular image in camera space remains highly challenging, particularly for encoding fine-grained metric depth geometry. To address this challenge, we propose a novel deep learning model that estimates the absolute hand position in camera space while enhancing the details of the reconstructed mesh. Specifically, our model employs a shared-weight feature encoder integrated with a depth regression head to extract latent hand representations and predict an initial coarse depth map. A key component is the Pseudo Stereo System, which generates pseudo-right features from left-view features and disparity cues, and establishes geometric constraints using a flexible feature-matching module. This design enables the model to learn depth-aware representations under training-time geometric supervision, while inference relies on a single RGB image. With this design, dense hand depth estimation is effectively guided by disparity map regression. Finally, a Transformer-based recovery module leverages 2D image-plane and depth features in combination to infer the 3D hand mesh. Extensive experiments on the FreiHAND dataset demonstrate that our model significantly outperforms existing methods in camera-centered 3D hand reconstruction and exhibits robust generalization across datasets in both camera-centered and root-relative settings. Our codes are publicly available at: https://github.com/ShaoXiang23/Pseudo-Stereo-Hand.