Research map / 3D vision / Gaussian splatting
GS-2DGS: Geometrically Supervised 2DGS for Reflective Object Reconstruction
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 GS-2DGS, a 2D Gaussian Splatting method for reflective object reconstruction that adds monocular normal and depth supervision from foundation models plus a physically based shading pipeline with deferred shading and a learned environment map.
Problem
Reconstructing highly reflective objects with Gaussian splatting yields noisy surfaces because view-dependent specular appearance violates multi-view consistency and Gaussians lack geometric constraints; SDF methods are accurate but slow.
Approach
Builds on 2DGS and supervises rendered normal and depth maps with pseudo ground truth from monocular foundation models (Marigold normals, Depth Pro depth), adds Cook-Torrance PBR shading with a trainable HDR environment map, and uses deferred shading instead of forward shading to better estimate lighting.
Key results
- Average Chamfer-L1 on Glossy Blender is 0.0068, best among Gaussian-based methods (R3DG 0.0303 avg reported in table for R3DG column; GShader 0.0553) and close to NeRO (0.0042), with training around 0.7 h versus 6 h for TensoSDF-type SDF methods as reported in the table.
- On Stanford-ORB, average Chamfer-L1 is 0.0047 versus 0.0070 for R3DG and 0.0139 for GS-IR.
- Relighting on Glossy Blender reaches average PSNR 19.56 and SSIM 0.8557 with 160 FPS, the highest PSNR among Gaussian methods (R3DG 19.19).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add geometric supervision from foundation models (ablation) | 2DGS · Glossy Blender | Chamfer-L1 distance: 0.0481 -> 0.0084 | Helped |
| add geometric supervision from foundation models (ablation) | 2DGS · Glossy Blender | PSNR: 26.23 -> 25.52 | Hurt |
| add PBR to geometric supervision (ablation) | 2DGS · Glossy Blender | Chamfer-L1 distance: 0.0084 -> 0.0074 | Helped |
| add PBR to geometric supervision (ablation) | 2DGS · Glossy Blender | PSNR: 25.52 -> 25.86 | Helped |
| add deferred shading to geometric supervision and PBR (ablation) | 2DGS · Glossy Blender | Chamfer-L1 distance: 0.0074 -> 0.0068 | Helped |
| add deferred shading to geometric supervision and PBR (ablation) | 2DGS · Glossy Blender | PSNR: 25.86 -> 26.76 | Helped |
Surprising findings
- Adding foundation-model geometric supervision alone raised Chamfer error from 0.0481 to 0.0084 in the ablation, while lowering rendering PSNR relative to the 2DGS baseline.
Limitations
- Still a slight performance gap versus SDF-based methods in reconstruction accuracy.
- Surfaces are less smooth than SDF-based results due to lack of an implicit smoothness prior in Gaussians.
- Relies on pseudo ground truth from monocular foundation models.
Experimental setup
Datasets: Glossy Blender (NeRO) Stanford-ORB
Benchmarks: Glossy Blender Stanford-ORB
Metrics: Chamfer-L1 distance PSNR SSIM FPS training time
Baselines: NeRO TensoSDF GShader GS-IR R3DG
Builds on: 2DGS 3DGS R3DG GS-IR Marigold Depth Pro NeRO
Closest papers by meaning
- Reflective Gaussian Splatting (ICLR 2025 · similarity 0.93)
- Anti-Aliased 2D Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.91)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- 3D-HGS: 3D Half-Gaussian Splatting (CVPR 2025 · similarity 0.91)
- SpecTRe-GS: Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
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.