Research map / 3D vision / NeRF and radiance fields
Structure from Collision
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.
Introduces structure from collision (SfC), estimating invisible internal structure of nontransparent objects from appearance changes during collision videos, via SfC-NeRF (PAC-NeRF-based with appearance-preserving, keyframe, mass constraints and volume annealing).
Problem
Estimating invisible internal structure (e.g., cavities) of nontransparent objects from multiview collision videos, which static neural 3D representations cannot recover because internal structure does not change static appearance.
Approach
Optimize the volume density grid of a physics-augmented NeRF over a collision video under physical (DiffMPM with known material and mass), appearance-preserving (first-frame pixel and depth losses with re-optimization of the external surface), and keyframe constraints, and escape local optima with volume annealing that alternately shrinks and expands the volume.
Key results
- SfC-NeRF achieves average chamfer distance 0.165 (x10^-3) versus 0.720 for Static and 0.656 for LPO across cavity sizes.
- Volume annealing ablation raises average CD from 0.165 to 0.316 (SfC-NeRF minus VA).
- Future prediction PSNR improves from 23.44 (PAC-NeRF) to 26.60 (SfC-NeRF).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| optimize internal structure with physical, appearance-preserving, and keyframe constraints plus volume annealing | SfC-NeRF · SfC dataset, cavity size average | CD (×103↓): 0.656 -> 0.165 | Helped |
| optimize internal structure with physical, appearance-preserving, and keyframe constraints plus volume annealing | SfC-NeRF · SfC dataset, cavity location average | CD (×103↓): 0.779 -> 0.281 | Helped |
| optimize the internal structure rather than keeping it filled | Future prediction · 14-frame test sequence | PSNR: 23.44 -> 26.60 | Helped |
| add mass loss (ablation) | SfC-NeRF · SfC dataset, cavity size average | CD (×103↓): 0.503 -> 0.165 | Helped |
| add appearance-preserving losses (ablation) | SfC-NeRF · SfC dataset, cavity size average | CD (×103↓): 0.688 -> 0.165 | Helped |
| add appearance-preserving training (ablation) | SfC-NeRF · SfC dataset, cavity size average | CD (×103↓): 0.335 -> 0.165 | Helped |
| add keyframe loss (ablation) | SfC-NeRF · SfC dataset, cavity size average | CD (×103↓): 0.186 -> 0.165 | Helped |
| add volume annealing (ablation) | SfC-NeRF · SfC dataset, cavity size average | CD (×103↓): 0.316 -> 0.165 | Helped |
| increase cavity size | SfC-NeRF · SfC dataset, cavity size | CD (×103↓): 0.081 -> 0.262 | Hurt |
| vary Young’s modulus | SfC-NeRF · SfC dataset, elastic materials | CD (×103↓): 0.289; 0.254; 0.195; 0.314; 0.374 | Mixed |
| vary Poisson’s ratio | SfC-NeRF · SfC dataset, elastic materials | CD (×103↓): 0.196; 0.198; 0.195; 0.207; 0.224 | Mixed |
Surprising findings
- Adding the mass loss to baseline GO/LPO did not improve performance and sometimes hurt it, attributed to lack of appearance-preserving mechanisms.
Limitations
- Assumes objects deform during collision, so rigid objects such as metals are hard to handle.
- Evaluated only on simulated data; real-data validation left for future work.
- Assumes physical properties, mass, camera parameters and collision objects are known.
Experimental setup
Datasets: SfC dataset (simulated with MLS-MPM) PAC-NeRF dataset
Benchmarks: SfC dataset
Metrics: chamfer distance anti-chamfer distance PSNR SSIM
Baselines: Static GO GOmass LPO LPOmass PAC-NeRF
Builds on: PAC-NeRF NeRF voxel-based NeRF (DVGO) material point method (DiffMPM) Lagrangian particle optimization
Closest papers by meaning
- PBR-NeRF: Inverse Rendering with Physics-Based Neural Fields (CVPR 2025 · similarity 0.88)
- Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis (NeurIPS 2024 · similarity 0.87)
- OMG: Opacity Matters in Material Modeling with Gaussian Splatting (ICLR 2025 · similarity 0.87)
- Splatter Image: Ultra-Fast Single-View 3D Reconstruction (CVPR 2024 · similarity 0.87)
- From Transparent to Opaque: Rethinking Neural Implicit Surfaces with $\alpha$-NeuS (NeurIPS 2024 · similarity 0.87)
- FruitNinja: 3D Object Interior Texture Generation with Gaussian Splatting (CVPR 2025 · similarity 0.86)
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.