Research map / Point clouds / Point cloud classification
GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting
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 GS-LiDAR, a panoramic Gaussian splatting framework with periodic-vibration 2D Gaussians and explicit ray-splat intersection that renders LiDAR range, intensity, and ray-drop maps for novel-view LiDAR synthesis on KITTI-360 and nuScenes.
Problem
Existing LiDAR novel view synthesis methods rely on NeRF, which is slow to train and render and is designed for symmetric scenes, making it ill-suited to driving scenarios with sparse LiDAR-only supervision.
Approach
Replaces NeRF with 2D Gaussian primitives carrying periodic vibration (temporal motion and opacity decay) and a panoramic rasterizer that computes exact ray-splat intersections for depth, while adding intensity and ray-drop SH coefficients refined by a U-Net.
Key results
- On KITTI-360 static scenes, depth RMSE drops from 3.2370 (LiDAR4D) to 2.8895 (10.7% reduction), and CD improves to 0.0847 from 0.0894.
- On nuScenes dynamic scenes, depth RMSE is 5.8925 vs 6.7831 for LiDAR4D (13.1% reduction) with Chamfer distance 0.2382 vs 0.2443.
- Training is 1.67x faster than LiDAR4D and rendering is reported as about 31x faster, reaching 11 FPS on a single RTX A6000 with 30k iterations taking about 1.5 hours.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ray-splat intersection (ablation) | GS-LiDAR · KITTI-360 | Depth RMSE: 6.6871 -> 3.1212 | Helped |
| add periodic vibration (ablation) | GS-LiDAR · KITTI-360 | Depth RMSE: 3.1738 -> 3.1212 | Helped |
| add median depth loss (ablation) | GS-LiDAR · KITTI-360 | Depth RMSE: 3.1414 -> 3.1212 | Helped |
| add depth distortion loss (ablation) | GS-LiDAR · KITTI-360 | Depth RMSE: 3.2016 -> 3.1212 | Helped |
| add normal consistency loss (ablation) | GS-LiDAR · KITTI-360 | Depth RMSE: 3.1993 -> 3.1212 | Helped |
| add chamfer distance loss (ablation) | GS-LiDAR · KITTI-360 | Depth RMSE: 3.1354 -> 3.1212 | Helped |
| add ray-drop refinement (ablation) | GS-LiDAR · KITTI-360 | Depth RMSE: 4.0791 -> 3.1212 | Helped |
Surprising findings
- Removing ray-splat intersection (using adapted 3DGS projection) collapses performance, e.g., point cloud F-score falls from 0.9231 to 0.5091 on KITTI-360.
- On nuScenes, LiDAR4D retains the best LPIPS and intensity SSIM/PSNR values for some metrics, so GS-LiDAR does not win on every metric.
Limitations
- Ray-drop modeling via Gaussians alone ignores distance effects, so a U-Net refinement is required; removing it degrades depth RMSE from 3.12 to 4.08.
- Evaluated only on three driving datasets with per-scene optimization; generalization beyond these is not explored.
- Rendering speed of 11 FPS is modest compared with the claim of fast rendering.
Experimental setup
Datasets: KITTI-360 nuScenes Waymo
Benchmarks: KITTI-360 nuScenes Waymo
Metrics: Chamfer distance F-score RMSE MedAE LPIPS SSIM PSNR
Baselines: LiDARsim NKSR PCGen LiDAR-NeRF LiDAR4D D-NeRF TiNeuVox-B K-Planes
Builds on: 3D Gaussian splatting 2D Gaussian splatting LiDAR-NeRF LiDAR4D NeRF-LiDAR
Closest papers by meaning
- NeRF-LiDAR: Generating Realistic LiDAR Point Clouds with Neural Radiance Fields (AAAI 2024 · similarity 0.89)
- LiDAR4D: Dynamic Neural Fields for Novel Space-time View LiDAR Synthesis (CVPR 2024 · similarity 0.89)
- LiDAR-RT: Gaussian-based Ray Tracing for Dynamic LiDAR Re-simulation (CVPR 2025 · similarity 0.89)
- SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving (CVPR 2025 · similarity 0.89)
- STGC-NeRF: Spatial-Temporal Geometric Consistency for LiDAR Neural Radiance Fields in Dynamic Scenes (AAAI 2025 · similarity 0.88)
- GeoNLF: Geometry guided Pose-Free Neural LiDAR Fields (NeurIPS 2024 · 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.