Research map / 3D vision / Gaussian splatting
TransmissiveGS: Residual-Guided Disentangled Gaussian Splatting for Transmissive Scene Reconstruction and Rendering
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 TransmissiveGS, a dual-Gaussian splatting framework that disentangles reflection and transmission in transmissive scenes using multi-view color residuals, a reflection light field, and a high-frequency loss.
Problem
Reconstructing and rendering transmissive scenes is hard because near-field reflections on the transmissive surface entangle with the transmitted content behind it, creating dual geometries and dual radiance per observation.
Approach
Model the scene with reflection Gaussians and scattering Gaussians composited by deferred shading; first fit multi-view consistent content with scattering Gaussians, then use the residual image as implicit geometric supervision for reflection Gaussian normals, and model near-field reflections with a position- and direction-conditioned reflection light field.
Key results
- Achieves average transmissive surface MAE of 7.46° versus 27.24° for the next best method (Ref-Gaussian) on seven synthetic scenes.
- Obtains best PSNR on all seven synthetic scenes and all four real-world scenes in the rendering comparison (e.g., 24.69 PSNR on facade ablation vs 24.52 without high-frequency loss).
- Ablation without dual-Gaussian drops PSNR to 22.98 on facade, the largest drop among the ablated components.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add dual-Gaussian representation (scattering Gaussians alongside reflection Gaussians) (ablation) | TransmissiveGS · facade | PSNR: 22.98 -> 24.69 | Helped |
| Replace near-field reflection light field with environment map reflection (ablation) | TransmissiveGS · facade | PSNR: 24.69 -> 24.20 | Hurt |
| Add decoupled optimization (freezing scattering attributes in Stage II and reflection geometry in Stage III) (ablation) | TransmissiveGS · facade | PSNR: 23.23 -> 24.69 | Helped |
| Add high-frequency loss (ablation) | TransmissiveGS · facade | PSNR: 24.52 -> 24.69 | Helped |
| Residual-guided reflection Gaussian geometry recovery with dual-Gaussian pipeline | TransmissiveGS vs prior GS methods · Blender synthetic transmissive surface (7 scenes avg) | MAE°: 7.46 vs best baseline Ref-Gaussian 27.24 | Helped |
| Full TransmissiveGS vs prior GS methods (3DGS, 2DGS, GaussianShader, Ref-Gaussian, Ref-GS, EnvGS) | photorealistic rendering · Blender synthetic and real-world scenes (PSNR) | PSNR: best in all 11 scenes, e.g. dome 21.22 vs 20.94 (Ref-Gaussian) | Helped |
Limitations
- Assumes both reflected and transmitted scene content are static.
- Reflection component degenerates when the surface is dominated by transmission with only faint reflections.
- Evaluates MAE only on synthetic scenes where ground-truth geometry is available.
Experimental setup
Datasets: NeRF-Casting dataset (compact, hatchback) custom Blender transmissive scenes self-captured real scenes (facade, glazing)
Benchmarks: TransmissiveGS Blender synthetic scenes NeRF-Casting compact and hatchback facade glazing
Metrics: PSNR SSIM LPIPS MAE°
Baselines: 3DGS 2DGS GaussianShader Ref-Gaussian Ref-GS EnvGS
Builds on: 2D Gaussian Splatting EnvGS Ref-GS Fourier feature encoding
Closest papers by meaning
- Reflective Gaussian Splatting (ICLR 2025 · similarity 0.90)
- Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis (NeurIPS 2024 · similarity 0.89)
- MaterialRefGS: Reflective Gaussian Splatting with Multi-view Consistent Material Inference (NeurIPS 2025 · similarity 0.89)
- Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.88)
- FatesGS: Fast and Accurate Sparse-View Surface Reconstruction Using Gaussian Splatting with Depth-Feature Consistency (AAAI 2025 · similarity 0.88)
- AtlasGS: Atlanta-world Guided Surface Reconstruction with Implicit Structured Gaussians (NeurIPS 2025 · similarity 0.88)
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.