Abstract
Synthetic data has emerged as a promising source for 3D human research as itoffers low-cost access to large-scale human datasets. To advance the diversityand annotation quality of human models, we introduce a new synthetic dataset,SynBody, with three appealing features: 1) a clothed parametric human modelthat can generate a diverse range of subjects; 2) the layered humanrepresentation that naturally offers high-quality 3D annotations to supportmultiple tasks; 3) a scalable system for producing realistic data to facilitatereal-world tasks. The dataset comprises 1.2M images with corresponding accurate3D annotations, covering 10,000 human body models, 1,187 actions, and variousviewpoints. The dataset includes two subsets for human pose and shapeestimation as well as human neural rendering. Extensive experiments on SynBodyindicate that it substantially enhances both SMPL and SMPL-X estimation.Furthermore, the incorporation of layered annotations offers a valuabletraining resource for investigating the Human Neural Radiance Fields (NeRF).