Date: November 2025.
Source: arXiv, Computer Science, Computer Vision and Pattern Recognition, preprint: 2511.05403.
Research Summary: PALM addresses a gap in hand modeling research, the lack of real-world datasets that combine accurate 3D geometry, high-resolution multi-view imagery, and a diverse subject population. The dataset includes 13,000 high-quality real-world 3dMD hand scans, registered against the MANO hand model, and 90,000 multi-view images from 263 subjects, capturing variation in skin tone, age, and hand geometry. The team also built PALM-Net, a baseline model that learns hand geometry and material properties through physically based inverse rendering, enabling realistic, relightable hand avatars personalized from a single image.
3dMD’s Role: PALM’s real-world 3D hand scans and multi-view RGB images were captured on a 3dMDhand7.t system using a 7-viewpoint, 21-camera array, reconstructed with 3dMD’s software-driven Active Stereo Photogrammetry technique. The dataset covers 263 subjects, 131 male and 132 female, with diverse skin tones and hand shapes, giving Meta Reality Labs and MPI-IS the real-world scan volume behind PALM-Net.
Article: PALM: A Dataset and Baseline for Learning Multi-subject Hand Prior.
Authors: Zicong Fan, Edoardo Remelli, David Dimond, Fadime Sener, Liuhao Ge, Bugra Tekin, Cem Keskin, Shreyas Hampali. Meta Reality Labs and Max Planck Institute for Intelligent Systems, Tubingen.