Research map / Point clouds / Registration and matching
GeoNLF: Geometry guided Pose-Free Neural LiDAR 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 GeoNLF, a pose-free neural LiDAR field that alternates bundle-adjusting NeRF optimization with graph-based geometric point cloud registration, plus selective reweighting and 3D geometric losses, for LiDAR novel view synthesis and multi-view registration on sparse sequences.
Problem
Novel view synthesis of LiDAR point clouds typically needs precomputed poses, while point cloud registration and prior pose-free NeRFs struggle to get globally accurate poses on sparse, low-frequency large-scale outdoor LiDAR sequences.
Approach
Alternate global bundle-adjusting neural LiDAR field optimization with a pure geometric graph-based optimizer (graph-based robust Chamfer distance), use selective reweighting that lowers the learning rate of neural fields for the top-k highest-loss frames so pose gradients still flow, and add 3D Chamfer and normal losses to the range-image supervision.
Key results
- On NuScenes, GeoNLF reaches CD 0.2408 and F-score 0.8647 versus 0.9691 and 0.8011 for the best baseline HASH-LN, and ATE 0.228 m versus 0.892 m for GeoTrans.
- On KITTI-360, GeoNLF reaches F-score 0.9178 and ATE 0.170 m, versus 0.8116 F-score and 0.246 m ATE for GeoTrans.
- Ablation on NuScenes shows removing the geo-optimizer raises ATE from 0.228 m to 1.328 m and CD from 0.2363 to 0.6180.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GeoNLF instead of the strongest listed baseline | LiDAR novel view synthesis · NuScenes | F-score: 0.8011 -> 0.8647 | Helped |
| Use GeoNLF instead of the strongest listed baseline | LiDAR novel view synthesis · KITTI-360 | F-score: 0.8116 -> 0.9178 | Helped |
| Use GeoNLF instead of the strongest listed baseline | pose estimation · NuScenes | ATE (m): 0.892 -> 0.228 | Helped |
| Use GeoNLF instead of the strongest listed baseline | pose estimation · KITTI-360 | ATE (m): 0.246 -> 0.170 | Helped |
| Add the Geo-optimizer (ablation) | GeoNLF · NuScenes | ATE (m): 1.328 -> 0.228 | Helped |
| Retain the control factor in graph-based RCD (ablation) | GeoNLF · NuScenes | ATE (m): 0.332 -> 0.228 | Helped |
| Add selective-reweighting (ablation) | GeoNLF · NuScenes | CD: 0.2654 -> 0.2363 | Helped |
| Add 3D geometric constraints (ablation) | GeoNLF · NuScenes | CD: 0.2877 -> 0.2363 | Helped |
| Use GeoNLF instead of the strongest listed baseline | LiDAR novel view synthesis · KITTI-360 | Depth PSNR: 25.6137 -> 25.2758 | Hurt |
Surprising findings
- Baseline registration methods (ICP, MICP, SGHR, HRegNet) fail completely on some sequences, producing large errors in averaged results.
- HRegNet trained on KITTI performs poorly on NuScenes, and it is sensitive to the temporal order of the point cloud sequence.
Limitations
- Not designed for dynamic scenes, which are common in autonomous driving.
- Relies on the temporal prior of the point cloud sequence, targeting point clouds within a single sequence.
- Learning-based registration baselines generalize poorly across datasets and require retraining, which limits fair comparison.
Experimental setup
Datasets: NuScenes KITTI-360 KITTI
Benchmarks: NuScenes KITTI-360
Metrics: CD F-score RMSE MedAE LPIPS SSIM PSNR RPE_t RPE_r ATE
Baselines: BARF-LN HASH-LN GeoTrans ICP MICP SGHR HRegNet
Builds on: LiDAR-NeRF BARF HASH LiDAR4D NeRFmm
Closest papers by meaning
- LiDAR4D: Dynamic Neural Fields for Novel Space-time View LiDAR Synthesis (CVPR 2024 · similarity 0.90)
- STGC-NeRF: Spatial-Temporal Geometric Consistency for LiDAR Neural Radiance Fields in Dynamic Scenes (AAAI 2025 · similarity 0.89)
- GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting (ICLR 2025 · similarity 0.88)
- AlignMiF: Geometry-Aligned Multimodal Implicit Field for LiDAR-Camera Joint Synthesis (CVPR 2024 · similarity 0.88)
- LiFi: LiDAR Generation from Multi-View Images via Geometric and Semantic Collaborative Guidance (NeurIPS 2026 · similarity 0.86)
- NeRF-LiDAR: Generating Realistic LiDAR Point Clouds with Neural Radiance Fields (AAAI 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.