Research map / 3D vision / Gaussian splatting
OMG: Opacity Matters in Material Modeling with Gaussian Splatting
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.
Adds a material-dependent cross-section MLP to 3DGS opacity (Beer-Lambert-inspired) so material parameters receive gradients from alpha, improving inverse rendering when plugged into three 3DGS baselines.
Problem
3D Gaussian Splatting-based inverse rendering treats opacity as a standalone parameter independent of material properties, leaving material estimation under-constrained.
Approach
Multiply each Gaussian's opacity by a learned cross-section term produced by a small MLP from material properties, with alpha = 1 - exp(-o G f(m)) following the Bouguer-Beer-Lambert law, so material gradients also come from opacity.
Key results
- On Synthetic4Relight, R3DG + Ours improves novel view synthesis PSNR from 34.07 to 34.47 and albedo PSNR from 28.70 to 29.31.
- On MIP-NeRF 360, GS-IR + Ours raises average PSNR from 25.38 to 25.85 and lowers average LPIPS from 0.265 to 0.254.
- GaussianShader + Ours improves average PSNR by about 0.3 dB on Shiny Blender (30.77 to 31.07) and 0.4 dB on Glossy Synthetic (27.25 to 27.68).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add OMG material-dependent opacity modeling | R3DG · Synthetic4Relight | PSNR (novel view synthesis): 34.07 -> 34.47 | Helped |
| Add OMG material-dependent opacity modeling | R3DG · Synthetic4Relight | MSE (roughness): 0.011 -> 0.007 | Helped |
| Add OMG material-dependent opacity modeling | GaussianShader · Shiny Blender | PSNR: 30.77 -> 31.07 | Helped |
| Add OMG material-dependent opacity modeling | GaussianShader · Glossy Synthetic | PSNR: 27.25 -> 27.68 | Helped |
| Add OMG material-dependent opacity modeling | GS-IR · MIP-NeRF 360 | PSNR (average): 25.38 -> 25.85 | Helped |
| Add OMG material-dependent opacity modeling | GS-IR · MIP-NeRF 360 | LPIPS (Flowers scene): 0.368 -> 0.401 | Hurt |
| Add OMG material-dependent opacity modeling | GS-IR · MIP-NeRF 360 | PSNR (Flowers scene): 20.43 -> 20.16 | Hurt |
Surprising findings
- Learned cross section is reported to assign lower values to windows and plastic bags, matching intuition, but the authors note it remains non-ideal in specular regions with strong lighting.
- Per-scene results show mixed outcomes; for example, on some scenes (Shiny Blender Car, Glossy Cat) the baseline is slightly better.
Limitations
- Cross section does not depend on light frequency (wavelength) in the current modeling.
- Learned cross section is not ideal in specular regions with strong lighting.
- Path length is set to a constant 1, an approximation justified heuristically.
Experimental setup
Datasets: Synthetic4Relight Shiny Blender Glossy Synthetic MIP-NeRF 360
Benchmarks: Synthetic4Relight Shiny Blender Glossy Synthetic MIP-NeRF 360
Metrics: PSNR SSIM LPIPS MSE MAE
Baselines: GaussianShader GS-IR R3DG
Builds on: 3D Gaussian Splatting Bouguer-Beer-Lambert law GaussianShader GS-IR R3DG
Closest papers by meaning
- MaterialRefGS: Reflective Gaussian Splatting with Multi-view Consistent Material Inference (NeurIPS 2025 · similarity 0.91)
- Subsurface Scattering for Gaussian Splatting (NeurIPS 2024 · similarity 0.90)
- Reflective Gaussian Splatting (ICLR 2025 · similarity 0.89)
- GS-IR: 3D Gaussian Splatting for Inverse Rendering (CVPR 2024 · similarity 0.88)
- Ref-GS: Directional Factorization for 2D Gaussian Splatting (CVPR 2025 · similarity 0.88)
- Splatter Image: Ultra-Fast Single-View 3D Reconstruction (CVPR 2024 · similarity 0.87)
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.