Research map / 3D vision / Gaussian splatting
Novel View Synthesis Under Large-Deviation Viewpoint for Autonomous Driving
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 an epipolar-attention illumination model and a planar-homography geometry augmentation to improve 3D Gaussian Splatting rendering for large-deviation viewpoints in driving scenes.
Problem
3D-GS novel view synthesis degrades when rendered viewpoints deviate far from the training trajectory in autonomous driving scenes, due to spherical-harmonic lighting overfitting and geometric inconsistency in texture-less road regions.
Approach
Replaces spherical harmonics with an epipolar cross-attention light-field decoder that aggregates colors from spatially nearest context views, and adds a plane-based supervision term that warps RGB and normal maps of detected planes to novel viewpoints via homography.
Key results
- On KITTI with translated viewpoints, the method reaches FID 0.0064 and KID 71.58 versus 3D-GS at 0.0104 and 79.98 (Table 1 column order as reported).
- Under a larger translation (Table 2), the method reaches KID 0.0138 on KITTI and FID 68.23 on Waymo, outperforming all baselines.
- Ablation on KITTI: the illumination model lowers FID from 98.24 to 77.24, and the full model reaches 71.58.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the epipolar-attention illumination model (ablation) | 3D-GS · KITTI, novel deviation viewpoints | FID: 98.24 -> 77.24 | Helped |
| add the epipolar-attention illumination model (ablation) | 3D-GS · KITTI, novel deviation viewpoints | KID: 0.0251 -> 0.0115 | Helped |
| add plane augmentation (ablation) | 3D-GS · KITTI, novel deviation viewpoints | FID: 98.24 -> 82.59 | Helped |
| add plane augmentation (ablation) | 3D-GS · KITTI, novel deviation viewpoints | KID: 0.0251 -> 0.0128 | Helped |
Surprising findings
- Scaffold-GS and GaussianPro perform worse than plain 3D-GS on large rotations, suggesting they overfit the training trajectory.
Limitations
- Evaluation uses only perceptual FID/KID since no ground truth exists for novel trajectory views.
- Evaluated on only 6 static scenes per dataset.
- Rendering speed is lower than the 3D-GS reference, and the paper reports no comparison against the base 3D-GS speed.
Experimental setup
Datasets: KITTI Waymo Open Dataset
Benchmarks: KITTI Waymo Open Dataset
Metrics: FID KID rendering speed FPS
Baselines: 3D-GS GaussianPro Scaffold-GS FSGS F2-NeRF
Builds on: 3D Gaussian Splatting Scaffold-GS PlaneRecTR GaussianPro
Closest papers by meaning
- GGS: Generalizable Gaussian Splatting for Lane Switching in Autonomous Driving (AAAI 2025 · similarity 0.89)
- Enhancing Novel View Synthesis via Geometry Grounded Set Diffusion (NeurIPS 2026 · similarity 0.88)
- Unifying Appearance Codes and Bilateral Grids for Driving Scene Gaussian Splatting (NeurIPS 2025 · similarity 0.88)
- DrivingGaussian: Composite Gaussian Splatting for Surrounding Dynamic Autonomous Driving Scenes (CVPR 2024 · similarity 0.87)
- SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving (CVPR 2025 · similarity 0.87)
- EVolSplat: Efficient Volume-based Gaussian Splatting for Urban View Synthesis (CVPR 2025 · 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.