Date: April 2019.
Source: arXiv.org, Computer Vision and Pattern Recognition.
Research Summary: Existing motion capture datasets are too small and varied to fully exploit deep learning, and each uses a different body parameterization, making them hard to combine. AMASS addresses this by unifying 15 optical marker-based mocap datasets into a single framework, using a new method, MoSh++, that converts sparse mocap markers into realistic, fully rigged 3D human meshes represented by the SMPL body model. MoSh++ recovers real dense-surface soft-tissue dynamics using DMPL and realistic hand pose using MANO. The result is the largest publicly available human motion database, more than 40 hours of motion, over 300 subjects, and more than 11,000 motions, all in a consistent, richly detailed format.
3dMD’s Role: AMASS’s own evaluation dataset, Synchronized Scans and Markers (SSM), was captured by synchronizing an OptiTrack marker system with a dynamic-3D/4D, dense-surface 3dMDbody System running at 60 frames per second, used to fine-tune and validate the accuracy of MoSh++ against real-world 3D scans. AMASS also inherits 3dMD-sourced data twice over: its soft-tissue dynamics come from DMPL, trained on the same 3dMD-captured Dyna dataset, and its 3D/4D hand motion comes from MANO, trained on 3dMD’s confirmed hand dataset.
Article: AMASS: Archive of Motion Capture as Surface Shapes.
Authors: Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, Michael J Black. Max Planck for Intelligent Systems, Perceiving Systems, Tuebingen, Germany.