PALM: A Dataset and Baseline for Learning Multi-subject Hand Prior. Z Fan, E Remelli, D Dimond, F Sener, L Ge, B Tekin, C Keskin, S Hampali.
The human hand is one of the most complex anatomical structures in nature… and one of the most demanding subjects to capture and reconstruct with near ground truth 3D-shape accuracy. Wearable devices, controllers, smart gloves, and gesture-based interfaces are not designed around a skeleton. They are designed around the actual surface of the hand they need to fit… across the full range of hand shapes, sizes, and proportions in the intended user population. Motion capture gloves record joint angles and finger kinematics, but the glove is on the hand during capture, physically altering the surface and making it impossible to capture the real 3D geometry of the hand or evaluate how a product prototype interacts with bare skin. The 3dMDhand system series was purpose-built to capture what no other method can… the actual dense-surface 3D geometry of the hand in motion.
Capturing 360-degree single or bimanual hand coverage dynamically in 3D/4D, the 3dMDhand delivers real-world dense-surface hand data at the anatomical fidelity that the most demanding applications require. No capture device worn on the hand. No instrument between the hand and the object of interaction. When a subject holds a device prototype or wears a product on their hand or wrist inside the 3dMD system, the capture is of the real interaction… the actual contact areas, gaps, and fit between the device and the hand surface as they exist in the real world. A glove-based system cannot provide this data because the glove itself is the capture instrument.
In healthcare, 3dMDhand data supports the assessment of hand mobility and function before and after surgery, tracks morphological change due to aging, arthritis, or injury, and provides the precise anatomical geometry needed for custom splint, orthotic, and prosthetic design. Beyond healthcare, before a humanoid robot hand can be teleoperated or trained, it needs to be designed… and achieving human-like dexterity requires understanding the actual 3D surface geometry, soft tissue deformation, and skin compliance of real human hands. Beyond design, dexterous humanoid robot hand training demands real grip dynamics, precision tool use, and bimanual coordination that sparse markers and 2D video cannot spatially reconstruct. Wearable tech and smart glove design require precise real-world 3D hand shape across diverse subject populations. Computer Vision and AI teams building hand gesture recognition systems, AR/XR UI control applications, and hand models need real-world 3D/4D hand sequences. The research team at the Max Planck Institute for Intelligent Systems, Perceiving Systems, in Tübingen developed MANO with 3dMD, a foundational hand model that became a cornerstone of computer vision research into human hand representation. Meta Reality Labs recently released PALM… 13k registered 3dMD hand scans from 263 subjects with 90k calibrated 3dMD multiview RGB images, each with corresponding MANO registrations. When ground truth hand data matters, it is captured on 3dMD.
Every 3dMD hand scan is a permanent, re-harvestable data asset. Dense-surface 3D meshes can be re-analyzed, re-measured, and re-harvested as research evolves, AI models develop, and new product questions emerge. The value of the data grows over time.
Whatever the application or use case, 3dMDhand systems capture and reconstruct it… in full 3D/4D dense-surface detail.
