Research map / 3D vision / Gaussian splatting
GGS: Generalizable Gaussian Splatting for Lane Switching in 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 GGS, a generalizable Gaussian splatting method for autonomous driving that uses virtual lane generation with lane switch-back cycles, a multi-lane diffusion loss, and MVS-based depth refinement to render novel views under large lateral viewpoint changes.
Problem
Novel view synthesis for lane switching in autonomous driving when training data comes only from a single lane, so large viewpoint changes lead to poor rendering of adjacent lanes.
Approach
Builds on MVSplat by generating virtual lanes via a lane converter and switching back to the real lane, creating a self-supervised lane-switch loss without multi-lane ground truth; adds a latent-diffusion-based multi-lane loss on virtual-lane images and a neighborhood-color, Metashape-MVS depth refinement.
Key results
- On KITTI Residential, GGS reports PSNR 26.26, SSIM 0.8840, LPIPS 0.0948 versus MVSplat's 22.45 / 0.6562 / 0.3008.
- Lane switching FID@LEFT 60.34 and FID@RIGHT 55.17, versus the best baseline READ at 79.79 and 76.87.
- Ablation: full model PSNR 29.12 vs baseline 25.14 on KITTI; removing the virtual lane module raises VGG from 210.8 to 257.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GGS instead of DC-Gaussian | KITTI Residential · KITTI Residential | PSNR: 25.63 -> 26.26 | Helped |
| Use GGS instead of READ | KITTI Road · KITTI Road | PSNR: 24.29 -> 25.01 | Helped |
| Use GGS instead of 3DGaussian | KITTI City · KITTI City | PSNR: 24.92 -> 26.79 | Helped |
| Use GGS instead of GaussianPro | Brno Urban Left side view · Brno Urban Left side view | PSNR: 25.75 -> 27.8 | Helped |
| Use GGS instead of DC-Gaussian | Brno Urban Left front side view · Brno Urban Left front side view | PSNR: 25.82 -> 25.85 | No clear effect |
| Use GGS instead of GaussianPro | Brno Urban Right side view · Brno Urban Right side view | PSNR: 28.27 -> 26.5 | Hurt |
| Use GGS instead of READ | KITTI lane switching · KITTI lane switching | FID: 79.79 -> 60.34 (LEFT); 76.87 -> 55.17 (RIGHT) | Helped |
| Add virtual lane generation module (ablation) | GGS · KITTI | PSNR: 26.76 -> 29.12 | Helped |
| Add multi-lane diffusion loss (ablation) | GGS · KITTI | PSNR: 28.97 -> 29.12 | Helped |
| Add depth refinement module (ablation) | GGS · KITTI | PSNR: 29.05 -> 29.12 | Helped |
Limitations
- Lane switching quality degrades in dynamic scenes with complex road conditions, multiple pedestrians, and mixed vehicles.
- Depends on Agisoft Metashape MVS point clouds as additional input and depth correction.
Experimental setup
Datasets: KITTI Brno Urban
Benchmarks: KITTI Brno Urban
Metrics: PSNR SSIM LPIPS VGG loss FID
Baselines: ADOP READ UC-NeRF 3DGaussian GaussianPro DC-Gaussian MVSplat
Builds on: MVSplat 3D Gaussian Splatting Stable Diffusion GaussianPro
Closest papers by meaning
- EGSRAL:An Enhanced 3D Gaussian Splatting Based Renderer with Automated Labeling for Large-Scale Driving Scene (AAAI 2025 · similarity 0.90)
- Novel View Synthesis Under Large-Deviation Viewpoint for Autonomous Driving (AAAI 2025 · similarity 0.89)
- SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving (CVPR 2025 · similarity 0.88)
- Unifying Appearance Codes and Bilateral Grids for Driving Scene Gaussian Splatting (NeurIPS 2025 · similarity 0.88)
- DisC-GS: Discontinuity-aware Gaussian Splatting (NeurIPS 2024 · similarity 0.87)
- DrivingGaussian: Composite Gaussian Splatting for Surrounding Dynamic Autonomous Driving Scenes (CVPR 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.