Accurate reconstruction of human hand geometry is essential for applications such as motion capture, virtual reality, and human-computer interaction. Beyond pose estimation, reliable reconstruction requires accurate subject-specific hand shape, particularly for tasks involving consistent surface correspondence such as video-to-texture mapping.
Recent work in hand reconstruction focuses on learning-based regression from RGB images, which achieves strong image-space alignment but does not necessarily recover metrically accurate, subject-specific shape parameters. This limits applicability in scenarios requiring geometrically consistent 3D representations, and there is currently no accessible solution for estimating MANO shape parameters directly from 3D scan data.
In this work, a scan-based approach for estimating subject-specific MANO shape parameters is proposed. High-resolution 3D hand scans are used as ground truth. The model is aligned using landmark-based registration and inverse kinematics, followed by shape optimization that minimizes surface distance. The method is evaluated using established mesh registration metrics and validated in a motion capture setting, including its application to video-to-texture mapping. Results indicate improved geometric consistency and reconstruction stability compared to image-based approaches, with reduced reconstruction errors in mocap experiments.
Overall, the proposed approach provides a metrically grounded solution for subject-specific hand shape estimation that is complementary to image-based methods. The method provides a reliable basis for applications requiring accurate individual hand geometry for the MANO model.
3D Hand Shape Estimation for Video-to-Texture Mapping
