Research map / 3D vision / Gaussian splatting
GIC: Gaussian-Informed Continuum for Physical Property Identification and Simulation
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 GIC, a hybrid 3D-Gaussian and MPM pipeline that reconstructs dynamic objects with motion-factorized Gaussians, builds a continuum via coarse-to-fine density filling, and uses both 3D surfaces and rendered masks to estimate physical properties.
Problem
Estimating physical properties (system identification) of deformable objects from multi-view video, where prior methods rely on known geometry, noisy reconstructions, or RGB rendering that distorts under large deformation.
Approach
Motion-factorized dynamic 3D Gaussians provide explicit shapes; a coarse-to-fine filling strategy converts Gaussian reconstructions into density fields and simulatable MPM continuum particles; these particles carry Gaussian opacity/scale attributes so they can render object masks as 2D shape guidance instead of RGB.
Key results
- On Spring-Gaus future-state simulation, mean CD 0.95 vs 5.16 for Spring-Gaus and mean PSNR 22.98 vs 17.06.
- On the PAC-NeRF dynamic reconstruction, mean CD 0.212 vs 0.234 for DefGS and 0.278 for PAC-NeRF.
- On D-NeRF, mean PSNR 40.85 vs 40.43 for DefGS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GIC dynamic Gaussian reconstruction instead of DefGS | Dynamic reconstruction · PAC-NeRF synthetic cross-shape, mean | CD ↓: 0.234 -> 0.212 | Helped |
| Use GIC dynamic Gaussian reconstruction instead of PAC-NeRF | Dynamic reconstruction · PAC-NeRF synthetic cross-shape, mean | EMD ↓: 0.026 -> 0.023 | Helped |
| Use GIC for future-state simulation instead of Spring-Gaus | Future-state simulation · Spring-Gaus synthetic future states, mean | CD ↓: 5.16 -> 0.95 | Helped |
| Use GIC for future-state simulation instead of PAC-NeRF | Future-state simulation · Spring-Gaus synthetic future states, mean | PSNR ↑: 15.77 -> 22.98 | Helped |
| Add 2D mask supervision (ablation) | GIC system identification · PAC-NeRF cross-shaped objects, Newtonian, log10(mu) | MAE: 2.19±2.90 -> 1.53±1.31 | Helped |
| Add 2D mask supervision (ablation) | GIC system identification · PAC-NeRF cross-shaped objects, Non-Newtonian, log10(mu) | MAE: 19.4±27.7 -> 13.5±18.2 | Helped |
| Add 2D mask supervision (ablation) | GIC system identification · PAC-NeRF cross-shaped objects, Plasticine, log10(tau_Y) | MAE: 9.04±2.37 -> 1.67±1.21 | Helped |
| Add 2D mask supervision (ablation) | GIC system identification · PAC-NeRF cross-shaped objects, Sand, theta_fric | MAE: 2.55±2.03 -> 4.18±0.52 | Hurt |
| Add 2D mask supervision (ablation) | GIC system identification · PAC-NeRF cross-shaped objects, Elasticity, log10(E) | MAE: 2.85±1.94 -> 2.43±3.29 | Helped |
Surprising findings
- Mask-only supervision (Ours*) outperforms PAC-NeRF on several cross-shaped object parameters, suggesting 2D silhouettes can be informative despite lacking appearance supervision.
Limitations
- Assumes continuum mechanics and known object constitutive model types
- Requires multi-view images with known camera poses
- Computationally expensive, at least 1.5 hours per object
Experimental setup
Datasets: PAC-NeRF synthetic dataset Spring-Gaus synthetic dataset Spring-Gaus real-world dataset D-NeRF
Benchmarks: PAC-NeRF synthetic dataset Spring-Gaus synthetic dataset D-NeRF
Metrics: Chamfer Distance Earth Mover's Distance PSNR SSIM Mean Absolute Error
Baselines: PAC-NeRF DefGS Spring-Gaus
Builds on: 3D Gaussian Splatting PAC-NeRF PhysGaussian DefGS Spring-Gaus
Closest papers by meaning
- OmniPhysGS: 3D Constitutive Gaussians for General Physics-Based Dynamics Generation (ICLR 2025 · similarity 0.85)
- PhysGaussian: Physics-Integrated 3D Gaussians for Generative Dynamics (CVPR 2024 · similarity 0.85)
- SGCR: Spherical Gaussians for Efficient 3D Curve Reconstruction (CVPR 2025 · similarity 0.84)
- FreeGave: 3D Physics Learning from Dynamic Videos by Gaussian Velocity (CVPR 2025 · similarity 0.84)
- GaussianPainter: Painting Point Cloud into 3D Gaussians with Normal Guidance (AAAI 2025 · similarity 0.84)
- Vid2Sim: Generalizable, Video-based Reconstruction of Appearance, Geometry and Physics for Mesh-free Simulation (CVPR 2025 · 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.