Review: SMPL: A Skinned Multi-Person Linear Model

This is my note on the ACM 2015 paper “SMPL: A Skinned Multi-Person Linear Model” by Loper, Matthew and Mahmood, Naureen and Romero, Javier and Pons-Moll, Gerard and Black, Michael J.

Introduction

SMPL is a skinned vertex-based model that accurately represents a wide variety of body shapes in natural human poses.

In SMPL, the pose information is independent from the shape information.

Some parameters are learned from data:

  • rest pose template
  • blend weights
  • pose-dependent blend shapes
  • identity-dependent blend shape
  • a regressor from vertices to joint locations

In SMPL, the authors learn blend shapes (identity, pose, soft-tissue dynamics) that are additively combined with a rest template before being tranformed by blend skinning.

The pose blend shapes are formulated as a linear function of the elements of the part rotation matrices.

Objective function penalizes the per-vertex disparities between registered meshses and the model.

Learn pose

Data: 1786 high-resolution 3D scans of different subjects in a wide variety of poses.

  1. align the template mesh to each scan
  2. optimize for:
    • the blend weights,
    • pose-dependent blend shapes
    • mean template shape (rest pose)
    • a regressor from shape to joint locations

Learn shape

Data: CAESAR dataset (~2000 scans per gender).

  1. Register a template mesh to each scan
  2. Pose normalize the data: which is critical when learning a vertex-based shape model.
  3. Use principal component analysis (PCA) to learn body shape blend shapes.