Research map / 3D vision / Gaussian splatting
DrivingGaussian: Composite Gaussian Splatting for Surrounding Dynamic Autonomous Driving Scenes
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 DrivingGaussian, a composite Gaussian splatting framework that models static backgrounds incrementally, dynamic objects via a Gaussian graph, and uses LiDAR priors to reconstruct surrounding dynamic driving scenes.
Problem
Reconstructing large-scale, surrounding multi-camera dynamic autonomous driving scenes with fast-moving objects and sparse views, where NeRF and prior 3DGS methods produce artifacts and blur.
Approach
Decomposes the scene into a progressively built static background (incremental static 3D Gaussians, bin by bin along the ego trajectory) and per-object dynamic Gaussian graph nodes placed by bounding box pose, combined into one composite Gaussian field and initialized/regularized with LiDAR priors.
Key results
- On nuScenes surrounding view reconstruction, Ours-L (LiDAR prior) reaches 28.74 PSNR / 0.865 SSIM / 0.237 LPIPS, versus 26.75 PSNR for EmerNeRF and 26.08 PSNR for 3DGS.
- On KITTI-360 monocular reconstruction, Ours-L reaches 25.62 PSNR / 0.868 SSIM, above the best baseline DNMP at 23.41 PSNR.
- Ablation: removing the Composite Dynamic Gaussian Graph drops PSNR from 28.36 to 26.97 (Ours-S comparison), and removing Incremental Static 3D Gaussians drops it to 27.72.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LiDAR prior initialization instead of SfM points (ablation) | DrivingGaussian · nuScenes | PSNR: 28.36 -> 28.74 | Helped |
| Use LiDAR prior initialization instead of SfM points (ablation) | DrivingGaussian · nuScenes | SSIM: 0.851 -> 0.865 | Helped |
| Use LiDAR prior initialization instead of SfM points (ablation) | DrivingGaussian · nuScenes | LPIPS: 0.256 -> 0.237 | Helped |
| Increase LiDAR initialization from 1M to 2M points (ablation) | DrivingGaussian · nuScenes | PSNR: 28.74 -> 28.78 | Helped |
| Add Incremental Static 3D Gaussians (ablation) | DrivingGaussian · nuScenes | PSNR: 27.72 -> 28.74 | Helped |
| Add Composite Dynamic Gaussian Graph (ablation) | DrivingGaussian · nuScenes | PSNR: 26.97 -> 28.74 | Helped |
| Add the TSSIM loss (ablation) | DrivingGaussian · nuScenes | PSNR: 27.88 -> 28.74 | Helped |
| Add the robust loss (ablation) | DrivingGaussian · nuScenes | PSNR: 28.05 -> 28.74 | Helped |
| Add the LiDAR loss (ablation) | DrivingGaussian · nuScenes | PSNR: 28.45 -> 28.74 | Helped |
| Use DrivingGaussian with LiDAR prior instead of EmerNeRF | nuScenes · nuScenes | PSNR: 26.75 -> 28.74 | Helped |
| Use DrivingGaussian with LiDAR prior instead of 3DGS | nuScenes · nuScenes | LPIPS: 0.298 -> 0.237 | Helped |
| Use DrivingGaussian with LiDAR prior instead of DNMP | KITTI-360 · KITTI-360 | PSNR: 23.41 -> 25.62 | Helped |
Surprising findings
- Random initialization with 600K or 1M points performs far worse (about 22.2 PSNR) than SfM initialization (28.36), showing initialization strongly matters for driving scenes.
- Increasing LiDAR point count does not monotonically improve results; overly dense points were argued to add redundant features that hurt optimization.
Limitations
- Relies on bounding boxes from datasets or a 3D detector/tracker and SAM for object extraction, so object quality depends on those components.
- Evaluated only on nuScenes and KITTI-360 with metrics PSNR/SSIM/LPIPS; no quantitative dynamic-object-specific evaluation in the main text.
Experimental setup
Datasets: nuScenes KITTI-360
Benchmarks: nuScenes KITTI-360
Metrics: PSNR SSIM LPIPS
Baselines: Instant-NGP NeRF+Time NSG Mip-NeRF Mip-NeRF360 Urban-NeRF S-NeRF SUDS EmerNeRF 3DGS 4DGS Point-NeRF
Builds on: 3D Gaussian Splatting 4DGS EmerNeRF COLMAP SAM
Closest papers by meaning
- DrivingForward: Feed-forward 3D Gaussian Splatting for Driving Scene Reconstruction from Flexible Surround-view Input (AAAI 2025 · similarity 0.92)
- DeSiRe-GS: 4D Street Gaussians for Static-Dynamic Decomposition and Surface Reconstruction for Urban Driving Scenes (CVPR 2025 · similarity 0.91)
- SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving (CVPR 2025 · similarity 0.90)
- DrivingRecon: Large 4D Gaussian Reconstruction Model For Autonomous Driving (NeurIPS 2025 · similarity 0.89)
- EGSRAL:An Enhanced 3D Gaussian Splatting Based Renderer with Automated Labeling for Large-Scale Driving Scene (AAAI 2025 · similarity 0.89)
- Unifying Appearance Codes and Bilateral Grids for Driving Scene Gaussian Splatting (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.