Research map / 3D vision / NeRF and radiance fields
STGC-NeRF: Spatial-Temporal Geometric Consistency for LiDAR Neural Radiance Fields in Dynamic 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 STGC-NeRF, a LiDAR NeRF that adds temporal (scene-flow-based) and spatial (neighborhood geometric/contextual feature) consistency regularizers to improve dynamic LiDAR scene novel view synthesis from posed scans.
Problem
Novel view synthesis of dynamic scenes from low-frequency, sparse LiDAR scans, where LiDAR NeRFs struggle with moving objects and sparse far-range points.
Approach
Uses an off-the-shelf pre-trained scene flow network (GMSF) to build temporal consistency losses between synthesized and real frames, and adds spatial losses matching eigen-based, height, and VLAD contextual descriptors of local point neighborhoods, on top of LiDAR4D as backbone.
Key results
- On KITTI-360 dynamic scenes, STGC-NeRF reaches CD 0.0997 vs 0.1089 for LiDAR4D (about 8.4% lower) and depth RMSE 3.0794 vs 3.5256.
- On nuScenes, CD 0.2204 vs 0.2443 for LiDAR4D (about 9.8% improvement).
- Ablation on KITTI-360 shows TGCR and SGCC together lower depth RMSE by about 10.3% and 10.2% respectively over the vanilla backbone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add spatial-temporal geometric consistency | LiDAR4D · KITTI-360 | CD: 0.1089 -> 0.0997 | Helped |
| add spatial-temporal geometric consistency | LiDAR4D · nuScenes | CD: 0.2443 -> 0.2204 | Helped |
| add LR1 temporal supervision (ablation) | LiDAR4D · KITTI-360 | depth RMSE: 3.5256 -> 3.2660 | Helped |
| add LR2 autoregressive-flow supervision (ablation) | LiDAR4D · KITTI-360 | depth RMSE: 3.5256 -> 3.2470 | Helped |
| add LR1 and LR2 temporal regularization (ablation) | LiDAR4D · KITTI-360 | depth RMSE: 3.5256 -> 3.1624 | Helped |
| add Fe geometric features (ablation) | LiDAR4D · KITTI-360 | depth RMSE: 3.5256 -> 3.2619 | Helped |
| add Fz height features (ablation) | LiDAR4D · KITTI-360 | depth RMSE: 3.5256 -> 3.3146 | Helped |
| add Fc contextual features (ablation) | LiDAR4D · KITTI-360 | depth RMSE: 3.5256 -> 3.2133 | Helped |
| add Fe, Fz, and Fc spatial constraints (ablation) | LiDAR4D · KITTI-360 | depth RMSE: 3.5256 -> 3.1644 | Helped |
| add LR1, LR2, Fe, Fz, and Fc (ablation) | LiDAR4D · KITTI-360 | depth RMSE: 3.5256 -> 3.0794 | Helped |
Limitations
- Depends on an off-the-shelf pretrained scene flow network (GMSF) whose errors can propagate into supervision.
- Gains are reported as relative percentages on a limited set of 6 KITTI-360 and 5 nuScenes dynamic scenes.
Experimental setup
Datasets: KITTI-360 nuScenes KITTI Scene Flow (for GMSF pretraining)
Benchmarks: KITTI-360 nuScenes KITTI-360 static scenes
Metrics: Chamfer distance F-score depth RMSE depth MedAE LPIPS SSIM PSNR intensity RMSE intensity MedAE
Baselines: LiDARsim NKSR PCGen LiDAR-NeRF D-NeRF TiNeuVox K-Planes LiDAR4D
Builds on: LiDAR4D LiDAR-NeRF GMSF NeRF VLAD
Closest papers by meaning
- LiDAR4D: Dynamic Neural Fields for Novel Space-time View LiDAR Synthesis (CVPR 2024 · similarity 0.92)
- LidaRF: Delving into Lidar for Neural Radiance Field on Street Scenes (CVPR 2024 · similarity 0.90)
- GeoNLF: Geometry guided Pose-Free Neural LiDAR Fields (NeurIPS 2024 · similarity 0.89)
- NeRF-LiDAR: Generating Realistic LiDAR Point Clouds with Neural Radiance Fields (AAAI 2024 · similarity 0.89)
- LiDAR-RT: Gaussian-based Ray Tracing for Dynamic LiDAR Re-simulation (CVPR 2025 · similarity 0.88)
- GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting (ICLR 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.