Research map / Point clouds / Point cloud classification
LiDAR4D: Dynamic Neural Fields for Novel Space-time View LiDAR Synthesis
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 LiDAR4D, a LiDAR-only dynamic neural field that uses a 4D hybrid planar-grid representation, scene flow priors from point clouds, and a U-Net ray-drop refinement for space-time LiDAR novel view synthesis.
Problem
Novel space-time view LiDAR synthesis for large-scale dynamic driving scenes is largely unexplored; prior LiDAR NVS methods are static and struggle with sparsity, scale, and ray-drop realism.
Approach
Combines low-resolution multi-planar (K-Planes-style) features with high-resolution 4D hash grids for large-scale scenes, adds flow-MLP motion priors supervised by chamfer distance between adjacent LiDAR frames, and refines ray-drop globally with a runtime-optimized U-Net over full depth/intensity/ray-drop maps.
Key results
- On KITTI-360, LiDAR4D reduces Chamfer Distance to 0.1089 vs 0.1438 for LiDAR-NeRF (24.3% reduction).
- On NuScenes, LiDAR4D achieves 24.2% lower CD than LiDAR-NeRF (0.2443 vs 0.3225) with best depth and intensity metrics among compared methods.
- On static KITTI-360 sequences, LiDAR4D matches or improves on LiDAR-NeRF (CD 0.0894 vs 0.0923), indicating no static-scene degradation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LiDAR4D instead of K-Planes | LiDAR NVS · KITTI-360 | CD: 0.1302 -> 0.1089 | Helped |
| Use LiDAR4D instead of K-Planes | LiDAR NVS · NuScenes | CD: 0.2982 -> 0.2443 | Helped |
| Use LiDAR4D instead of LiDAR-NeRF | LiDAR NVS · KITTI-360 Static Scene Sequence | CD: 0.0923 -> 0.0894 | Helped |
Limitations
- Long-distance vehicle motion and occlusion remain open; dynamic object reconstruction lags static quality.
- Foreground and background are hard to separate.
- Quantitative NVS evaluation is limited to the ego-car trajectory and does not decouple novel spatial and temporal views.
Experimental setup
Datasets: KITTI-360 NuScenes
Benchmarks: KITTI-360 NuScenes
Metrics: Chamfer Distance F-score RMSE MedAE LPIPS SSIM PSNR
Baselines: LiDARsim NKSR PCGen LiDAR-NeRF D-NeRF TiNeuVox K-Planes
Builds on: K-Planes Instant-NGP LiDAR-NeRF tiny-cuda-nn
Closest papers by meaning
- STGC-NeRF: Spatial-Temporal Geometric Consistency for LiDAR Neural Radiance Fields in Dynamic Scenes (AAAI 2025 · similarity 0.92)
- GeoNLF: Geometry guided Pose-Free Neural LiDAR Fields (NeurIPS 2024 · similarity 0.90)
- 3D LiDAR Mapping in Dynamic Environments using a 4D Implicit Neural Representation (CVPR 2024 · similarity 0.90)
- GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting (ICLR 2025 · similarity 0.89)
- NeRF-LiDAR: Generating Realistic LiDAR Point Clouds with Neural Radiance Fields (AAAI 2024 · similarity 0.88)
- Dynamic LiDAR Re-simulation using Compositional Neural Fields (CVPR 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.