Research map / Point clouds / Point cloud classification
L2RSI: Cross-view LiDAR-based Place Recognition for Large-scale Urban Scenes via Remote Sensing Imagery
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.
Builds LiRSI-XA dataset and L2RSI, which aligns LiDAR BEV submaps with remote-sensing semantic submaps via contrastive learning and refines retrieval with Gaussian-mixture spatio-temporal particle estimation (STPE) for large-scale cross-view LiDAR place recognition.
Problem
LiDAR-based place recognition normally needs costly prior 3D maps; using overhead remote sensing imagery as the reference creates cross-view and cross-modal gaps, and single queries are ambiguous over a 100 km^2 area.
Approach
Map LiDAR submaps and remote-sensing submaps into a shared semantic space (segmentation labels, MAE-based dual-branch encoder with symmetric InfoNCE), then aggregate retrieval results from a sequence of queries as a mixture of Gaussians propagated by odometry, instead of particle filtering that discards particles.
Key results
- On LiRSI-XA Test Set-G (100 km^2, 63k RSI database), L2RSI (w. STPE) reaches Recall@1 of 83.27% and Recall@10 of 87.82% at a 30 m threshold, versus 22.95% for Sample4Geo.
- Without any fine-tuning, L2RSI (w. STPE) reaches Recall@1 above 40% on both LiRSI-Oxford trajectories, while baselines such as LIP-Loc and GeoDTR stay near 1% Recall@1.
- STPE adds about 31.7 ms per query (sampling rate 30%), and the total of 37.5 ms allows real-time operation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use L2RSI with STPE instead of Sample4Geo | LiRSI-XA and LiRSI-Oxford place recognition · LiRSI-XA Test Set-G (100km²) | Recall@1 (%): 22.95 -> 83.27 | Helped |
| Add semantic segmentation (ablation) | L2RSI · LiRSI-XA Test Set-S | Recall@1 (%): 50.43 -> 88.93 | Helped |
| Add semantic segmentation (ablation) | L2RSI · LiRSI-Oxford 11-14-02-26 | Recall@1 (%): 8.13 -> 42.29 | Helped |
| Add STPE (ablation) | L2RSI · LiRSI-XA Test Set-S | Recall@1 (%): 30.05 -> 88.93 | Helped |
| Add STPE (ablation) | L2RSI · LiRSI-Oxford 11-14-02-26 | Recall@1 (%): 10.19 -> 42.29 | Helped |
| Add orientation input (ablation) | L2RSI · LiRSI-XA Test Set-S | Recall@1 (%): 54.37 -> 88.93 | Helped |
| Add orientation input (ablation) | L2RSI · LiRSI-Oxford 11-14-02-26 | Recall@1 (%): 4.92 -> 42.29 | Helped |
| Use MAE-B rather than the other evaluated semantic encoders (ablation) | L2RSI · LiRSI-XA Test Set-S | Recall@1 (%): 73.77; 59.38; 72.58; 83.75; 84.72 -> 88.93 | Helped |
| Set STPE sampling rate λ to 30% rather than the other tested rates (ablation) | L2RSI · LiRSI-XA Test Set-S | Recall@1 (%): 85.98; 88.07; 88.93; 88.44; 88.44; 88.07; 88.07; 87.82; 88.07; 88.44 | Mixed |
| Increase yaw and translation noise in the STPE motion model (ablation) | L2RSI STPE · LiRSI-XA Test Set-G (100km²) | Recall@1 (%): 83.27(-0.00) -> 35.91(-47.36) | Hurt |
Surprising findings
- Replacing particle filtering with the proposed aggregation is needed because particles decay under unreliable single queries; PF reached only 47.85% Recall@1 on Test Set-G versus 83.27% for STPE.
- DINOv2-S gave strong results on Test Set-S and Test Set-G but collapsed to 1.40% Recall@1 on the Oxford trajectory, unlike MAE-B.
Limitations
- Requires an external magnetometer for directional reference (about ±10 degrees accuracy).
- Designed for urban scenes; performance may degrade in rural areas with sparse semantics.
- Temporal gap of three years between LiDAR and remote sensing imagery causes failures due to changed structures.
Experimental setup
Datasets: LiRSI-XA LiRSI-Oxford LiRSI-Kitti360 SemanticKITTI KITTI-360 Oxford Radar RobotCar
Benchmarks: LiRSI-XA LiRSI-Oxford LiRSI-Kitti360
Metrics: Recall@1 Recall@10 Recall@30 runtime (ms) parameters (M)
Baselines: LIP-Loc GeoDTR Sample4Geo FRGeo SuperGlobal particle filtering
Builds on: MAE InfoNCE SphereFormer FastGICP SuperGlobal
Closest papers by meaning
- HOTFormerLoc: Hierarchical Octree Transformer for Versatile Lidar Place Recognition Across Ground and Aerial Views (CVPR 2025 · similarity 0.88)
- ForestLPR: LiDAR Place Recognition in Forests Attentioning Multiple BEV Density Images (CVPR 2025 · similarity 0.87)
- LiSA: LiDAR Localization with Semantic Awareness (CVPR 2024 · similarity 0.86)
- Multi-Modal Aerial-Ground Cross-View Place Recognition with Neural ODEs (CVPR 2025 · similarity 0.85)
- From Coarse to Fine: A Matching and Alignment Framework for Unsupervised Cross-View Geo-Localization (AAAI 2025 · similarity 0.85)
- LightLoc: Learning Outdoor LiDAR Localization at Light Speed (CVPR 2025 · similarity 0.85)
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.