Date: November 2022.
Source: Computer Vision – ECCV 2022. Lecture Notes in Computer Science, vol 13662. Springer, Cham. https://doi.org/10.1007/978-3-031-20086-1_33.
Research Summary: Statistical 3D shape models of the head, hands, and full body are widely used in computer vision and graphics, but existing models of the head and hands fail to capture the full range of motion for these body parts. Feet have received even less attention, despite being crucial to modeling human movement, with applications in biomechanics, animation, and footwear. The problem is that previous body part models are trained on 3D scans isolated to individual parts, which cannot capture how those parts actually move relative to the rest of the body, such as how the head moves relative to the neck. SUPR (Sparse Unified Part-Based Representation) addresses this by jointly training a full-body model and specific part models on a combined dataset of full-body and body-part scans, an expressive head model (SUPR-Head), an articulated hand model (SUPR-Hand), and a novel foot model (SUPR-Foot), each with vertices strictly influenced by a sparse set of nearby joints. SUPR-Foot uses real-world 4D scans of feet to train an extended kinematic tree capturing toe motion, along with a non-linear deformation function that predicts how the foot deforms under ground contact, based on foot pose, shape, and contact conditions. SUPR is trained on an unprecedented 1.2 million real-world body, head, hand, and foot scans, and it generalizes better than existing expressive body models and body-part models while capturing each body part’s full range of motion.
3dMD’s Role: SUPR’s full-body training data of 775,481 scans capturing a wide range of body types and poses, was captured using a 60fps 3dMDbody22.u System built with 22 Modular Camera Units (MCUs), each with a pair of stereo cameras, a color camera, and a speckle-light projector, outputting meshes of approximately 150,000 vertices. A separate, purpose-built 3dMDfoot10.t System, using 10 pairs of stereo cameras including dedicated soles-of-the-foot cameras, captured 69,257 scans through a transparent glass platform to model how the foot deforms under ground contact, contributing to SUPR-Foot, the first publicly available articulated model of the human foot.
Article: SUPR: A Sparse Unified Part-Based Human Representation.
Authors: Ahmed AA Osman, Timo Bolkart, Dimitrios Tzionas, Michael J Black. Max Planck Institute for Intelligent Systems, Perceiving Systems, Tübingen, Germany.