Research map / 3D vision / Depth and stereo
LightLoc: Learning Outdoor LiDAR Localization at Light Speed
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.
LightLoc speeds up scene-coordinate-regression LiDAR localization to about 1 hour of training per new scene by freezing a pretrained backbone, adding classification-guided features, and pruning well-learned samples during training.
Problem
Scene coordinate regression for outdoor LiDAR localization is accurate but takes days of training per new scene because of large coverage areas and huge numbers of frames.
Approach
Freeze a scene-agnostic backbone trained on many scenes, guide scene-specific regression heads with features from a 5-minute K-means sample-classification network, and drop well-converged samples using the variance of median loss (redundant sample downsampling).
Key results
- On QEOxford, LightLoc reaches 0.83 m / 1.12° average error with 1 hour of training versus 53 hours for LiSA (0.95 m).
- On Oxford, LightLoc achieves 2.67 m / 1.25°, the best position error among compared methods, with 1 hour of training.
- Ablation on QEOxford: SCG cuts training time by about 45.8% and RSD reduces training from 2.4h to 2h with negligible accuracy loss; LOAM+LightLoc SLAM reduces mean error from 87.81 m to 4.48 m.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace LiSA with LightLoc | LiDAR localization · QEOxford | mean position error: 0.95 -> 0.83 | Helped |
| replace LiSA with LightLoc | LiDAR localization · Oxford | mean position error: 2.74 -> 2.67 | Helped |
| replace LiSA with LightLoc | LiDAR localization · NCLT | mean orientation error: 2.31 -> 2.80 | Hurt |
| add a scene-agnostic backbone and freeze it while training scene-specific prediction heads (ablation) | LightLoc · QEOxford | mean position/orientation error: 1.02/1.23 vs 1.02/1.33 | No clear effect |
| add sample classification guidance (SCG) (ablation) | LightLoc · QEOxford | mean position/orientation error: 1.02/1.33 -> 0.82/1.10 | Helped |
| add redundant sample downsampling (RSD) (ablation) | LightLoc without SCG · QEOxford | mean position/orientation error: 1.02/1.33 -> 1.03/1.36 | No clear effect |
| add RSD while retaining SCG (ablation) | LightLoc · QEOxford | training time: 1.3h -> 1h | Helped |
| vary SCG cluster count k1 (ablation) | LightLoc · QEOxford | mean position/orientation error: 0.90/1.18; 0.85/1.12; 0.83/1.12; 0.89/1.17; 0.89/1.16 | Mixed |
| replace RSD with random sampling (ablation) | LightLoc · QEOxford | mean position/orientation error: RSD 0.82/1.11 vs random 1.02/1.30 at 15%; RSD 0.83/1.12 vs random 1.12/1.37 at 25%; RSD 0.93/1.15 vs random 1.25/1.55 at | Hurt |
| replace RSD with uniform sampling (ablation) | LightLoc · QEOxford | mean position/orientation error: 0.82/1.13 at 15%; 0.92/1.14 at 25%; 1.25/1.52 at 35% | Mixed |
Surprising findings
- RSD with a 25% pruning ratio showed no detectable accuracy drop, while uniform or random sampling degraded at higher ratios.
Limitations
- Regression-based methods cannot generalize to unknown regions; NCLT results discard localization-failure areas.
- Accuracy is not best on all NCLT trajectories and orientation accuracy trails DiffLoc on QEOxford.
- Backbone is trained on only 18 nuScenes scenes, so generalization to other scenes/LiDAR types is untested.
Experimental setup
Datasets: nuScenes Oxford Radar RobotCar QEOxford NCLT
Benchmarks: QEOxford Oxford Radar RobotCar NCLT
Metrics: mean position error (m) mean orientation error (deg) training time (h) runtime (ms)
Baselines: LiSA SGLoc DiffLoc HypLiLoc NIDALoc STCLoc PointLoc PosePN PosePN++ PoseMinkLoc PoseSOE
Builds on: ACE SGLoc GLACE LOAM
Closest papers by meaning
- DiffLoc: Diffusion Model for Outdoor LiDAR Localization (CVPR 2024 · similarity 0.89)
- SOAR: Regression-based LiDAR Relocalization for UAVs (NeurIPS 2026 · similarity 0.88)
- LiSA: LiDAR Localization with Semantic Awareness (CVPR 2024 · similarity 0.87)
- GBlobs: Explicit Local Structure via Gaussian Blobs for Improved Cross-Domain LiDAR-based 3D Object Detection (CVPR 2025 · similarity 0.87)
- ForestLPR: LiDAR Place Recognition in Forests Attentioning Multiple BEV Density Images (CVPR 2025 · similarity 0.87)
- HOTFormerLoc: Hierarchical Octree Transformer for Versatile Lidar Place Recognition Across Ground and Aerial Views (CVPR 2025 · 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.