SMPL

A parametric 3D human body model that generates a full body mesh from a small set of shape and pose parameters.

SMPL was published in 2015 by Michael Black’s group at the Max Planck Institute for Intelligent Systems in Germany, learned from thousands of body scans. Given shape parameters β (usually 10 numbers controlling height and build) and pose parameters θ (rotations for 23 joints plus global orientation), it generates a triangle mesh with 6,890 vertices; linear blend skinning plus corrective deformations keep the mesh looking natural as joints bend. SMPL-X, released in 2019, folds in the MANO hand model and the FLAME face model for 10,475 vertices total. SMPL has become the common format for human motion data: AMASS unifies 15 motion-capture datasets into SMPL parameters, and humanoid robots doing motion retargeting usually start from SMPL motion clips. The model is free for research use only; commercial use requires a license.

ExampleH2O first optimizes SMPL’s shape parameters so the body skeleton proportions match Unitree’s H1 humanoid, then retargets about 10,000 SMPL motion clips from AMASS onto the H1 to train a whole-body motion-tracking policy.

Also called

Skinned Multi-Person Linear Model, SMPL-X, SMPL+H

Related

MANO · AMASS (Archive of Motion Capture as Surface Shapes) · Motion Retargeting · Human Mesh Recovery · Human Pose Estimation · H2O

Sources

SMPL 官方项目页（Max Planck Institute for Intelligent Systems） (Chinese)
Expressive Body Capture: 3D Hands, Face, and Body from a Single Image（SMPL-X，CVPR 2019）
Learning Human-to-Humanoid Real-Time Whole-Body Teleoperation（H2O）