Research map / Probabilistic and scientific modeling / Dynamical systems
Bayesian Differentiable Physics for Cloth Digitalization
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 a Bayesian differentiable cloth model that infers heterogeneous, stochastic stretching and bending parameters of real fabrics from a single Cusick drape silhouette, plus a new Cusick drape dataset.
Problem
Digitalizing specific real cloth samples (replicating their physical behavior) from very small, accurately measured Cusick drape test data, where existing methods assume homogeneous materials, deterministic dynamics, or need large datasets or 3D motion.
Approach
Treat cloth physical parameters as latent variables with a learnable Gaussian variational posterior, optimized via variational inference (reparameterization) through a differentiable implicit-Euler cloth simulator with per-element (heterogeneous) stiffness, and a differentiable renderer that compares predicted silhouettes with the single observed drape image.
Key results
- BDP achieves average MSE 3.84e-2 and Hausdorff distance 2.17e-2 m, versus 5.72e-2 / 3.40e-2 for HOMO and 5.51e-2 / 3.28e-2 for HETER.
- In the optimizer comparison, BDP reaches average MSE 4.52e-2 and H. Dis 2.01e-2, versus REMBO (6.03e-2 / 4.25e-2) and HeSBO (5.90e-2 / 4.35e-2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the Bayesian differentiable cloth model instead of HOMO | cloth digitalization · Cusick drape | Avg MSE: 5.72 × 10−2 -> 3.84 × 10−2 | Helped |
| Use the Bayesian differentiable cloth model instead of HOMO | cloth digitalization · Cusick drape | Avg H. Dis: 3.40 × 10−2 -> 2.17 × 10−2 | Helped |
| Use the Bayesian differentiable cloth model instead of HETER | cloth digitalization · Cusick drape | Avg MSE: 5.51 × 10−2 -> 3.84 × 10−2 | Helped |
| Use the Bayesian differentiable cloth model instead of HETER | cloth digitalization · Cusick drape | Avg H. Dis: 3.28 × 10−2 -> 2.17 × 10−2 | Helped |
| Add element-wise material heterogeneity | HOMO cloth model · Cusick drape | Avg MSE: 5.72 × 10−2 -> 5.51 × 10−2 | Helped |
| Add element-wise material heterogeneity | HOMO cloth model · Cusick drape | Avg H. Dis: 3.40 × 10−2 -> 3.28 × 10−2 | Helped |
| Use gradient-based optimization instead of REMBO | BDP cloth model · Cusick drape | Avg MSE: 6.03 × 10−2 -> 4.52 × 10−2 | Helped |
| Use gradient-based optimization instead of REMBO | BDP cloth model · Cusick drape | Avg H. Dis: 4.25 × 10−2 -> 2.01 × 10−2 | Helped |
| Use gradient-based optimization instead of HeSBO | BDP cloth model · Cusick drape | Avg MSE: 5.90 × 10−2 -> 4.52 × 10−2 | Helped |
| Use gradient-based optimization instead of HeSBO | BDP cloth model · Cusick drape | Avg H. Dis: 4.35 × 10−2 -> 2.01 × 10−2 | Helped |
Surprising findings
- The heterogeneous deterministic model (HETER) fits the training sample better than homogeneous HOMO, but still fails to generalize to unseen samples of the same cloth type.
Limitations
- Requires prior knowledge of the underlying physics; cannot plug-and-play on arbitrary data.
- Models only static and dynamic draping; does not yet model buckling or other dynamics stochasticity.
- Garment results are not compared against real garment motion capture.
Experimental setup
Datasets: Cusick drape silhouettes collected under BS EN ISO 9073-9:2008
Benchmarks: Cusick drape silhouettes (train on 1, test on remaining 11 per cloth type)
Metrics: MSE Hausdorff distance radius-angle graph
Baselines: HOMO HETER REMBO HeSBO
Builds on: differentiable cloth simulation variational inference DIB-R differentiable renderer implicit Euler method
Closest papers by meaning
- Learning 3D Garment Animation from Trajectories of A Piece of Cloth (NeurIPS 2024 · similarity 0.85)
- NeuralClothSim: Neural Deformation Fields Meet the Thin Shell Theory (NeurIPS 2024 · similarity 0.84)
- PGC: Physics-Based Gaussian Cloth from a Single Pose (CVPR 2025 · similarity 0.84)
- Garment Recovery with Shape and Deformation Priors (CVPR 2024 · similarity 0.83)
- DeClotH: Decomposable 3D Cloth and Human Body Reconstruction from a Single Image (CVPR 2025 · similarity 0.83)
- Reconstruction of Manipulated Garment with Guided Deformation Prior (NeurIPS 2024 · similarity 0.83)
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.