Research map / 3D vision / Face and shape
ShapeBoost: Boosting Human Shape Estimation with Part-Based Parameterization and Clothing-Preserving Augmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ShapeBoost, a monocular human shape recovery framework that parameterizes body shape by bone lengths and part-slice widths and augments training data with clothing-preserving image transformations.
Problem
Monocular human pose and shape recovery overfits on body shape due to scarce diverse training data, failing on extreme body shapes and on people in thick clothing or occlusion.
Approach
Replace PCA shape coefficients with a part-based parameterization (bone lengths plus mean widths of part slices) reconstructed by a semi-analytical algorithm (analytical stretching/broadening followed by an MLP correction), and augment data by segmenting the human, reshaping it with affine transforms, and pasting it back onto an inpainted background so clothing is preserved.
Key results
- On SSP-3D, ShapeBoost (SMPL) reaches 11.4 mm PVE-T-SC, versus 12.2 mm for retrained SHAPY and 13.3 mm for the best prior method listed.
- On HBW validation, ShapeBoost (SMPL) achieves the best hip circumference error (42 mm) and best height error (58 mm), while P2P-20K is 25 mm versus SHAPY's 21 mm on the test set.
- Ablation: replacing the beta parameterization with the part-based one lowers PVE-T-SC from 12.3 to 11.4 mm; removing augmentation raises it to 12.1 mm.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use part-based parameterization instead of direct PCA shape parameters (ablation) | ShapeBoost · SSP-3D test | PVE-T-SC (mm): 12.3 -> 11.4 | Helped |
| use part-based parameterization instead of direct PCA shape parameters (ablation) | ShapeBoost · HBW validation | P2P20K error (mm): 26.0 -> 25.1 | Helped |
| increase part slicing number from n=1 to n=2 (ablation) | ShapeBoost image-based shape estimation · SSP-3D test | PVE-T-SC (mm): 11.4 -> 11.6 | Hurt |
| increase part slicing number from n=1 to n=2 (ablation) | ShapeBoost image-based shape estimation · HBW validation | P2P20K error (mm): 25.1 -> 26.2 | Hurt |
| use semi-analytical reconstruction instead of analytical-only reconstruction (ablation) | ShapeBoost shape reconstruction · synthetic shape reconstruction, 0% noise | V2V error (mm): 6.14 -> 0.69 | Helped |
| use semi-analytical reconstruction instead of neural-network-only reconstruction (ablation) | ShapeBoost shape reconstruction · synthetic shape reconstruction, 0% noise | V2V error (mm): 1.82 -> 0.69 | Helped |
| increase part slicing number from n=1 to n=3 (ablation) | ShapeBoost shape reconstruction · synthetic shape reconstruction, 5% noise | V2V error (mm): 8.83 -> 7.63 | Helped |
| remove data augmentation (ablation) | ShapeBoost · SSP-3D test | PVE-T-SC (mm): 11.4 -> 12.1 | Hurt |
| remove data augmentation (ablation) | ShapeBoost · HBW validation | P2P20K error (mm): 25.1 -> 26.5 | Hurt |
| remove shape decomposition in addition to data augmentation (ablation) | ShapeBoost · SSP-3D test | PVE-T-SC (mm): 12.1 -> 12.4 | Hurt |
| remove shape decomposition in addition to data augmentation (ablation) | ShapeBoost · HBW validation | P2P20K error (mm): 26.5 -> 27.0 | Hurt |
| add Model Agency Dataset to training | ShapeBoost · HBW validation | P2P20K error: drops by less than 1 point | Helped |
Surprising findings
- Using a larger part slicing number (n=2) did not improve image-based shape estimation over n=1, which the authors attribute to harder CNN regression.
- On HBW test, ShapeBoost is not best overall: SHAPY has lower height and P2P-20K errors, so the claimed superiority is partial.
Limitations
- Pose estimation fails under severe occlusion.
- SMPL's shape space limits accuracy for children and muscular bodies.
- Pose-dependent shape deformation is ignored in predicting bone lengths and part widths.
Experimental setup
Datasets: 3DPW Human3.6M COCO AGORA Model Agency Dataset AMASS
Benchmarks: SSP-3D HBW
Metrics: PVE-T-SC P2P-20K height error chest error waist error hip circumference error MPJPE PA-MPJPE V2V error
Baselines: HMR SPIN HybrIK SHAPY CLIFF LVD Sengupta et al. 2021a PARE
Builds on: SMPL HybrIK SHAPY Instaboost
Closest papers by meaning
- HumanPro: Single-view 3D Clothed Human Reconstruction with Progressive Normal Guidance (AAAI 2026 · similarity 0.88)
- Garment Recovery with Shape and Deformation Priors (CVPR 2024 · similarity 0.86)
- MultiGO: Towards Multi-level Geometry Learning for Monocular 3D Textured Human Reconstruction (CVPR 2025 · similarity 0.86)
- VS: Reconstructing Clothed 3D Human from Single Image via Vertex Shift (CVPR 2024 · similarity 0.85)
- AiOS: All-in-One-Stage Expressive Human Pose and Shape Estimation (CVPR 2024 · similarity 0.85)
- HiLo: Detailed and Robust 3D Clothed Human Reconstruction with High-and Low-Frequency Information of Parametric Models (CVPR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.