Research map / 3D vision / 3D generation
Dynamic LiDAR Re-simulation using Compositional Neural Fields
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 DyNFL, an SDF-based neural field method that decomposes driving scenes into static and per-vehicle neural fields and composes them via a ray drop test to re-simulate and edit LiDAR scans.
Problem
High-fidelity LiDAR re-simulation and editing (novel views, object insertion/removal, trajectory changes) in dynamic driving scenes, where prior neural LiDAR methods are limited to static scenes or lack physical sensing effects.
Approach
Combines SDF-based volume rendering with a LiDAR physical model (two-way transmittance, ray drop) and composes independently rendered static and dynamic neural fields per ray, selecting the nearest non-dropped range instead of blending sampled points across fields.
Key results
- On Waymo Dynamic, DyNFL reaches MedAE 3.0 cm and CD 10.9 cm versus UniSim's 6.1 cm and 14.3 cm, and LiDARsim's 11.5 cm and 31.1 cm.
- Compositing by per-field rendering and ray drop test roughly halves MedAE across all points compared with UniSim-style point blending.
- On static scenes, DyNFL achieves the lowest MAE on TownClean (26.7 cm) and TownReal (33.9 cm) among compared methods, including NFL (32.0 and 39.2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DyNFL instead of UniSim | LiDAR NVS · Waymo Dynamic | MAE: 35.6 -> 30.8 | Helped |
| use DyNFL instead of UniSim | LiDAR NVS · Waymo Dynamic | MedAE Dyn: 14.3 -> 8.5 | Helped |
| use DyNFL instead of NFL | LiDAR NVS · TownClean | MAE: 32.0 -> 26.7 | Helped |
| use DyNFL instead of NFL | LiDAR NVS · TownReal | MAE: 39.2 -> 33.9 | Helped |
| use DyNFL instead of i-NGP | LiDAR NVS · Waymo interp | MAE: 26.4 -> 28.3 | Hurt |
| use DyNFL instead of NFL | LiDAR NVS · Waymo NVS | MAE: 32.6 -> 28.6 | Helped |
| replace active-sensor SDF volume rendering with passive-sensor SDF volume rendering (ablation) | DyNFL · TownClean | MAE: -1.5 | Hurt |
| replace active-sensor SDF volume rendering with passive-sensor SDF volume rendering (ablation) | DyNFL · Waymo Interp | MAE: 0.1 | No clear effect |
| replace active-sensor SDF volume rendering with passive-sensor SDF volume rendering (ablation) | DyNFL · Waymo Dynamic | MAE: -0.3 | No clear effect |
| remove surface-point SDF regularization (ablation) | DyNFL · TownReal | MAE: -3.3 | Hurt |
| remove surface-point SDF regularization (ablation) | DyNFL · Waymo Interp | MAE: -0.3 | Hurt |
Surprising findings
- UniSim-style point blending across fields performs notably worse than per-field rendering with ray drop composition, with MedAE roughly doubled.
Limitations
- Difficulty synthesizing moving vehicles from unseen view angles.
- Relies on accurate bounding boxes and trajectories of moving objects.
- Baselines LiDARsim and UniSim were re-implemented since they are not open-sourced.
Experimental setup
Datasets: Waymo Open Waymo Dynamic TownClean TownReal Waymo Interp Waymo NVS
Benchmarks: Waymo Dynamic TownClean TownReal Waymo Interp Waymo NVS
Metrics: MAE MedAE Chamfer distance MedAE Dyn Intensity RMSE detection agreement recall precision IoU
Baselines: LiDARsim UniSim NFL i-NGP DS-NeRF URF
Builds on: Neural LiDAR Fields NeuS Neural Scene Graph UniSim Instant-NGP
Closest papers by meaning
- LiDAR-RT: Gaussian-based Ray Tracing for Dynamic LiDAR Re-simulation (CVPR 2025 · similarity 0.89)
- LiDAR4D: Dynamic Neural Fields for Novel Space-time View LiDAR Synthesis (CVPR 2024 · similarity 0.88)
- STGC-NeRF: Spatial-Temporal Geometric Consistency for LiDAR Neural Radiance Fields in Dynamic Scenes (AAAI 2025 · similarity 0.86)
- 3D LiDAR Mapping in Dynamic Environments using a 4D Implicit Neural Representation (CVPR 2024 · similarity 0.86)
- NeRF-LiDAR: Generating Realistic LiDAR Point Clouds with Neural Radiance Fields (AAAI 2024 · similarity 0.86)
- LidaRF: Delving into Lidar for Neural Radiance Field on Street Scenes (CVPR 2024 · similarity 0.86)
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.