Research map / 3D vision / Gaussian splatting
DC-Gaussian: Improving 3D Gaussian Splatting for Reflective Dash Cam Videos
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 DC-Gaussian, a 3D Gaussian Splatting method that jointly models windshield reflections and occlusions in dash cam videos via learned opacity maps, illumination-aware obstruction modeling, and geometry-guided Gaussian enhancement.
Problem
Novel view synthesis from dash cam videos is degraded by windshield obstructions (reflections, occlusions, stains) that violate multi-view consistency, causing 3DGS to model them as static geometry.
Approach
Decomposes each frame into transmission and obstruction using a learnable per-view 2D opacity map; models the globally shared obstruction appearance with a multi-resolution hash encoding conditioned on camera position via Latent Intensity Modulation gates; and enhances 3DGS geometry with MVS depth priors filtered on obstruction-suppressed images.
Key results
- On BDD100K, DC-Gaussian reaches 29.44 PSNR / 0.914 SSIM / 0.143 LPIPS versus 28.02 PSNR for 3DGS.
- On DCVR, DC-Gaussian reaches 24.74 PSNR / 0.822 SSIM / 0.202 LPIPS, versus 24.41 PSNR for Zip-NeRF and 23.73 for 3DGS.
- Ablation on DCVR obstruction areas: full model reaches 26.30 PSNR versus 23.99 for the bare 3DGS baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add global-shared hash encoding via NOM (ablation) | 3DGS · DCVR obstruction-influenced areas | PSNR: 23.99 -> 25.21 | Helped |
| Add adaptive image decomposition (AD) (ablation) | DC-Gaussian with NOM · DCVR obstruction-influenced areas | PSNR: 25.21 -> 25.65 | Helped |
| Add Latent Intensity Modulation (LIM) (ablation) | DC-Gaussian with NOM and AD · DCVR obstruction-influenced areas | PSNR: 25.65 -> 25.90 | Helped |
| Add Geometry-Guided Gaussian Enhancement (G3E) (ablation) | DC-Gaussian with NOM, AD, and LIM · DCVR obstruction-influenced areas | PSNR: 25.90 -> 26.30 | Helped |
| Vary the obstruction-inpainting threshold τ among 0.3, 0.5, and 0.7 (ablation) | DC-Gaussian · DCVR | PSNR: 26.27, 26.30, 26.28 | No clear effect |
| Preprocess images with DSRNet reflection removal before training 3DGS (ablation) | 3DGS · BDD100K | PSNR: 28.02 -> 27.99 | No clear effect |
| Preprocess images with DSRNet reflection removal before training 3DGS (ablation) | 3DGS · DCVR | PSNR: 23.73 -> 23.72 | No clear effect |
| Replace 3DGS with DC-Gaussian | 3DGS · BDD100K | PSNR: 28.02 -> 29.44 | Helped |
| Replace ZipNeRF with DC-Gaussian | ZipNeRF · DCVR | PSNR: 24.41 -> 24.74 | Helped |
Surprising findings
- DSRNet-based reflection removal before 3DGS training does not improve novel view synthesis (23.72 vs 23.73 PSNR on DCVR).
- Standard NeRF-in-the-wild fails to separate obstructions from transient content.
Limitations
- Evaluated only on single-sequence videos.
- Not specifically designed for dynamic scenes; performance drops for fast-moving dynamic objects.
- Opacity loss assumes opaque objects occupy only small portions of the windshield.
Experimental setup
Datasets: BDD100K DCVR
Benchmarks: BDD100K DCVR
Metrics: PSNR SSIM LPIPS FPS
Baselines: 3DGS Zip-NeRF NeRF-W GaussianPro DSRNet NIR NeRFRen
Builds on: 3D Gaussian Splatting NeRF-in-the-wild multi-resolution hash encoding PatchMatchNet UC-NeRF
Closest papers by meaning
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.88)
- COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · similarity 0.88)
- DrivingForward: Feed-forward 3D Gaussian Splatting for Driving Scene Reconstruction from Flexible Surround-view Input (AAAI 2025 · similarity 0.88)
- DrivingGaussian: Composite Gaussian Splatting for Surrounding Dynamic Autonomous Driving Scenes (CVPR 2024 · similarity 0.87)
- EVolSplat: Efficient Volume-based Gaussian Splatting for Urban View Synthesis (CVPR 2025 · similarity 0.87)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 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.