Research map / 3D vision / Gaussian splatting
RestorGS: Depth-aware Gaussian Splatting for Efficient 3D Scene Restoration
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 RestorGS, a depth-aware 3D Gaussian Splatting framework that decouples clear and degraded Gaussians via spherical-harmonic appearance decoupling and depth-guided physical degradation modeling to restore underwater, nighttime, and hazy 3D scenes.
Problem
Novel view synthesis and 3D reconstruction from degraded scenes (underwater, nighttime, hazy) with multi-view inconsistency, where 3DGS relies on high-quality clean inputs.
Approach
Decouple each Gaussian into a clear version (color-mapped spherical harmonics via a color MLP) and a degraded version (learnable degradation factors as SH), then model degradation with a depth-derived attenuation map and a CNN-predicted illumination map under a unified physical scattering model, trained self-supervised without clear reference views.
Key results
- On SeaThru-NeRF, average PSNR improves by 0.45 dB over WaterSplatting across four underwater scenes.
- On Gaussian-DK nighttime dataset, average PSNR 25.04 vs 24.70 for Gaussian-DK and 18.55 for 3DGS.
- Ablation on the Mip-NeRF360 hazy setting: full model reaches PSNR 29.19, SSIM 0.928, LPIPS 0.050, versus 28.46 PSNR without appearance decoupling.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add appearance decoupling (ablation) | RestorGS · Ablation study | PSNR: 28.46 -> 29.19 | Helped |
| add color MLP (ablation) | RestorGS · Ablation study | PSNR: 28.83 -> 29.19 | Helped |
| use physical modeling instead of simple summation (ablation) | RestorGS · Ablation study | PSNR: 28.62 -> 29.19 | Helped |
| add depth guidance (ablation) | RestorGS · Ablation study | PSNR: 29.05 -> 29.19 | Helped |
| use RestorGS rather than Gaussian-DK | 3D scene restoration · Gaussian-DK dataset average | PSNR: 24.70 -> 25.04 | Helped |
| use RestorGS rather than WaterSplatting | 3D scene restoration · Curacao | PSNR: 29.36 -> 31.95 | Helped |
| use RestorGS rather than WaterSplatting | 3D scene restoration · Red Sea | PSNR: 32.51 -> 30.79 | Hurt |
| use RestorGS rather than Gaussian-DK | 3D scene restoration · Garden | PSNR: 27.74 -> 28.32 | Helped |
Limitations
- Quantitative evaluation uses reconstruction of degraded views because clear ground truth is unavailable in real scenes.
- Hazy evaluation relies on synthetically generated haze from a modified 3DGS depth estimate.
- Slight inferiority to WaterSplatting in the Panama underwater scene.
Experimental setup
Datasets: SeaThru-NeRF Gaussian-DK Mip-NeRF360
Benchmarks: SeaThru-NeRF Gaussian-DK Mip-NeRF360
Metrics: PSNR SSIM LPIPS
Baselines: SeaThru-NeRF ZipNeRF HDR-NeRF NeRF-W 3DGS Gaussian-DK WaterSplatting
Builds on: 3D Gaussian Splatting NeRF atmospheric scattering model
Closest papers by meaning
- VastGaussian: Vast 3D Gaussians for Large Scene Reconstruction (CVPR 2024 · similarity 0.90)
- ReorgGS: Equivalent Distribution Reorganization for 3D Gaussian Splatting (NeurIPS 2026 · similarity 0.89)
- DepthSplat: Connecting Gaussian Splatting and Depth (CVPR 2025 · similarity 0.89)
- SplatFormer: Point Transformer for Robust 3D Gaussian Splatting (ICLR 2025 · similarity 0.89)
- PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
- Generative Sparse-View Gaussian Splatting (CVPR 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.