← ArchivePaper2020
STAR: Sparse Trained Articulated Human Body Regressor
STAR learns spatially local pose-corrective blend shapes with sparse joint influence, reducing SMPL parameters by 80% while improving deformation realism.
Abstract
The SMPL body model is widely used for the estimation, synthesis, and analysis of 3D human pose and shape. While popular, we show that SMPL has several limitations and introduce STAR, which is quantitatively and qualitatively superior to SMPL. First, SMPL has a huge number of parameters resulting from its use of global blend shapes. These dense pose-corrective offsets relate every vertex on the mesh to all the joints in the kinematic tree, capturing spurious long-range correlations. To address this, we define per-joint pose correctives and learn the subset of mesh vertices that are influenced by each joint movement. This sparse formulation results in more realistic deformations and significantly reduces the number of model parameters to 20% of SMPL. When trained on the same data as SMPL, STAR generalizes better despite having many fewer parameters. Second, SMPL factors pose-dependent deformations from body shape while, in reality, people with different shapes deform differently. Consequently, we learn shape-dependent pose-corrective blend shapes that depend on both body pose and BMI. Third, we show that the shape space of SMPL is not rich enough to capture the variation in the human population. We address this by training STAR with an additional 10,000 scans of male and female subjects, and show that this results in better model generalization.
How to read this
- Category
- Method: a sparse, articulated parametric human body model
- Contributions
- STAR, a body model with per-joint pose-corrective blend shapes and learned sparse joint-to-vertex influence, replacing SMPL's dense global blend shapes and cutting parameters substantially
- Shape-dependent pose-corrective blend shapes that depend on both pose and BMI, since differently shaped people deform differently
- A richer shape space than SMPL to better capture human body variation
- Context
- A direct re-formulation and critique of SMPL (Loper et al.), keeping the skinned-linear paradigm but localizing the corrective deformations.Builds on: SMPL: A Skinned Multi-Person Linear Model
- Correctness
- Reports being quantitatively and qualitatively superior to SMPL when trained on the same data, attributing it to removing spurious long-range vertex-to-joint correlations; a reader should note the improvements are relative to SMPL on shared training data, so the comparison is the right frame rather than an absolute ceiling.
- Clarity
- Accessible, especially with SMPL background; a first pass conveys the sparsity argument, a second pass covers the corrective-blendshape formulation.
- How to read it
- First pass on the three SMPL limitations and how STAR addresses each; second pass on the sparse per-joint correctives and the pose-plus-BMI shape dependence if you build on parametric bodies.
Builds on
Related work
- SUPR: A Sparse Unified Part-Based Human Representation 2022 / Eurographics
- SCANimate: Weakly Supervised Learning of Skinned Clothed Avatar Networks 2021 / CVPR
- Dyna: A Model of Dynamic Human Shape in Motion 2015 / SIGGRAPH
- Animatable Neural Radiance Fields for Modeling Dynamic Human Bodies 2021 / CVPR
Keywords
This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →