Research map / 3D vision / Gaussian splatting
SpecTRe-GS: Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D 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.
Introduces SpecTRe-GS, a 3D Gaussian Splatting method that separately models highly specular and rough surfaces and traces secondary rays in 3D Gaussian space to render inter-reflections of nearby objects.
Problem
3DGS struggles to model high-frequency view-dependent appearance and global illumination (inter-reflections) on highly specular surfaces reflecting nearby objects, since low-order SH and baked incident lighting cannot capture them and rasterization cannot efficiently trace secondary rays.
Approach
Model highly specular surfaces with ideal mirror reflection whose incident radiance is obtained by tracing one secondary ray through the 3D Gaussian point cloud (OptiX with stretched icosahedron proxies), while rough surfaces keep per-Gaussian SH; geometry is refined with StableNormal normal priors and numerical gradients of ray-traced incident radiance.
Key results
- On synthetic scenes, SpecTRe-GS reaches 23.28 PSNR in reflective regions versus 21.98 for the next-best 3iGS and 21.57 for 3DGS-DR (entire image 29.41 PSNR).
- On real-world scenes, reflective-region PSNR is 22.77 versus 21.31 for 3DGS.
- Rendering is only 50% to 100% slower than a pure rasterizer on the test scenes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SpecTRe-GS instead of the strongest baseline, 3iGS | SpecTRe-GS · Synthetic (reflective regions) | PSNR: 21.98 -> 23.28 | Helped |
| Use SpecTRe-GS instead of the strongest baseline, 3DGS | SpecTRe-GS · Real-world (reflective regions) | PSNR: 21.31 -> 22.77 | Helped |
| Add normal prior guidance (ablation) | SpecTRe-GS · Synthetic scenes (reflective regions) | PSNR: 19.80 -> 24.05 | Helped |
| Add joint geometry optimization from incident lighting (ablation) | SpecTRe-GS · Synthetic scenes (reflective regions) | PSNR: 19.52 -> 24.05 | Helped |
| Add ray-traced indirect incident lighting (ablation) | SpecTRe-GS · Synthetic scenes (reflective regions) | PSNR: 22.19 -> 24.05 | Helped |
| Add progressive learning (ablation) | SpecTRe-GS · Synthetic scenes (reflective regions) | PSNR: 23.16 -> 24.05 | Helped |
| Add depth-aware ray perturbation (ablation) | SpecTRe-GS · Synthetic scenes (reflective regions) | PSNR: 23.88 -> 24.05 | Helped |
Surprising findings
- Normal prior guidance degrades GOF, which relies solely on SH color, because it suppresses high-frequency appearance through false geometry.
Limitations
- Only models specular inter-reflections from rough surfaces; avoids multi-bounce path tracing.
- Traces only one secondary ray per pixel in reflective regions and approximates specular reflection as ideal mirror.
- Monocular normal prior may fail under strong semantic ambiguity (e.g., mirrors with hidden borders), causing geometry initialization failure.
Experimental setup
Datasets: 6 custom synthetic Blender Cycles scenes 2 real-world captured scenes (REALBOWL, REALPOT)
Benchmarks: custom synthetic scenes (HELMET, MARBLETABLE, VASE, POT, TOASTER, MIRROR) custom real-world scenes (REALBOWL, REALPOT)
Metrics: PSNR SSIM LPIPS
Baselines: 3DGS GOF 3iGS GShader 3DGS-DR
Builds on: 3D Gaussian Splatting GOF 3DGRT 3DGS-DR GShader
Closest papers by meaning
- SpectroMotion: Dynamic 3D Reconstruction of Specular Scenes (CVPR 2025 · similarity 0.91)
- Spec-Gaussian: Anisotropic View-Dependent Appearance for 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.91)
- GS-2DGS: Geometrically Supervised 2DGS for Reflective Object Reconstruction (CVPR 2025 · similarity 0.91)
- Reflective Gaussian Splatting (ICLR 2025 · similarity 0.90)
- Normal-GS: 3D Gaussian Splatting with Normal-Involved Rendering (NeurIPS 2024 · similarity 0.90)
- MaterialRefGS: Reflective Gaussian Splatting with Multi-view Consistent Material Inference (NeurIPS 2025 · similarity 0.89)
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.