Date: November 2015.
Source: ACM Transactions on Graphics (TOG), Proceedings of ACM SIGGRAPH Asia 2015, Vol. 34, No. 6, Article 248.
Research Summary: SMPL (Skinned Multi-Person Linear model) is a learned, parametric model of human body shape and pose-dependent shape variation, more accurate than previous models and compatible with existing graphics pipelines. It is a skinned, vertex-based model that accurately represents a wide variety of body shapes across natural human poses. The model’s parameters, including the rest pose template, blend weights, pose-dependent blend shapes, identity-dependent blend shapes, and a joint regressor, are all learned from data, made possible by a formulation where pose-dependent blend shapes are a linear function of the pose rotation matrices. This formulation enables training the entire model from a large number of aligned 3D body scans of different people in different poses. SMPL outperforms a Blend-SCAPE model trained on the same data, and its dynamic extension, DMPL, further extends the model to realistically capture soft-tissue deformation.
3dMD’s Role: SMPL’s pose-dependent blend shapes, which help determine how the body deforms with natural movement, were trained on a multi-pose dataset of 1,786 real-world scans across 40 subjects, captured on MPI’s static 3dMDbody scanner. This body pose dataset supplemented SMPL’s separate static shape space trained on the CAESAR body scan database. CAESAR, which consists of one standing A-pose and two seated poses, was designed for measurement use in human factors and ergonomics. With MPI-IS’s adoption of the 3dMDbody22.u system capturing at 60fps, the Dyna model was trained on the 40,000-scan, 10-subject real-world dataset. As such, SMPL was then extended into a new model, DMPL (Dynamic-SMPL), which adds soft-tissue deformation, trained on the same real-world dataset as Dyna.
Article: SMPL: A Skinned Multi-Person Linear Model.
Authors: Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, Michael J Black. Max Planck Institute for Intelligent Systems, Tübingen, Germany.