cs.CVAug 15, 2026
SaveLightLoc++: Sensor-Robust Representation Learning for Efficient Outdoor LiDAR Localization
Organizations: Fujian Key Laboratory of Urban Intelligent Sensing and Computing, and with Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, School of Informatics, Xiamen University, 422 Siming Road South, Xiamen, FJ 361005, P.R. China · School of Engineering Mathematics and Technology, University of Bristol, Bristol BS8 1TH, United Kingdom · School of Computer Science and Engineering, Northeastern University, Shenyang 110819, China
Abstract
Scene coordinate regression (SCR) achieves strong performance in outdoor LiDAR localization, but it usually requires scene-specific training that can take days, limiting practical deployment. Recent works improve training efficiency by decoupling SCR into a scene-agnostic backbone and scene-specific prediction heads, where the backbone is pretrained on source datasets and frozen for new scenes, and only lightweight heads are optimized. However, we find that this paradigm heavily depends on the pretrained backbone. Existing decoupled methods can match conventional SCR methods fully optimized for each new scene when LiDAR configurations are similar to those used during backbone pretraining, but their accuracy drops noticeably on datasets collected with different LiDAR sensors. This suggests that efficient LiDAR localization requires representations that capture stable scene geometry across LiDAR configurations. Motivated by this observation, we propose LightLoc++, a sensor-robust and efficient outdoor LiDAR localization framework. To support sensor-robust representation learning, we introduce SULID, a synchronized urban multi-LiDAR dataset with representative 32-, 64-, and 128-beam rotating LiDARs, extensive cross-sensor overlap, and diverse urban scenes. Using SULID, we pretrain a sensor-robust backbone through cross-sensor consistency learning. LightLoc++ further preserves efficient new-scene learning by incorporating sample classification guidance and redundant sample downsampling, which reduce regression ambiguity and computational redundancy in large-scale outdoor scenes. Extensive experiments on multiple outdoor LiDAR localization benchmarks demonstrate that LightLoc++ achieves state-of-the-art localization performance with the lowest new-scene training cost among compared methods. Code and dataset will be made available at https://github.com/liw95/LightLoc-PlusPlus.
Figures & tables
Fig. 1: LiDAR localization performance vs. training efficiency. The figure compares normalized localization accuracy and training efficiency across multiple LiDAR benchmarks collected with different sensors [ 3 , 4 , 5 , 6 , 7 ] . Larger values indicate better performance. LightLoc++ achieves consistently strong localization performance across diverse sensing modalities while maintaining high training efficiency.
| Datasets | Reference | Diverse LiDAR | Cross-LiDAR Consistency | Diverse Scene | Scene Types | Total Distance |
| KITTI [ 63 ] | CVPR’12 | ✗ | ✗ | ✓ | Open | 44 Km |
| NCLT [ 64 ] | IJRR’16 | ✗ | ✗ | ✗ | Structured | 147 Km |
| Oxford [ 5 ] | ICRA’20 | ✗ | ✗ | ✗ | Open | 280 Km |
| MulRan [ 4 ] | ICRA’20 | ✗ | ✗ | ✓ | Open, Structured | 123 Km |
| Apollo-SouthBay [ 22 ] | CVPR’19 | ✗ | ✗ | ✓ | Open | 381 Km |
| nuScenes [ 23 ] | CVPR’20 | ✗ | ✗ | ✓ | Open | 242 Km |
TABLE I: Comparison of LiDAR localization datasets. LiDAR diversity refers to the use of sensors with different configurations (e.g., beam numbers and sampling densities), while cross-sensor consistency indicates aligned observations of the same scene across sensors. Scene types are categorized as open (sparse urban environments with weak structural constraints), structured (regular layouts with repeated geometric patterns), dense (cluttered environments with high structural density and strong occlusions), and degenerate (low-feature or geometrically ambiguous environments with limited observability).
Fig. 2: Overview of the SULID dataset. (a) Aerial views of the collected scenes, including Campus-A , Campus-B , Urban-Road , and Urban-Hub , with LiDAR trajectories overlaid, covering diverse urban environments. (b) The data collection platform is equipped with multiple LiDAR sensors of different configurations, along with their spatial arrangement and coordinate systems, enabling synchronized multi-LiDAR observations.
| Sensor | Resolution (V H) | FOV (H V) | Range | FPS |
| Ouster OS1-128 | 128 2048 | 360 ∘ 45 ∘ | 120m | 10Hz |
| Ouster OS1-64 | 64 1024 | 360 ∘ 45 ∘ | 120m | 10Hz |
| Hesai XT-32 | 32 1800 | 360 ∘ 31 ∘ | 120m | 10Hz |
TABLE II: Specifications of multi-LiDAR sensors used in SULID.
Fig. 3: Cross-sensor point-cloud observations in SULID. Point clouds captured simultaneously at the same location by Ouster OS1-128, Ouster OS1-64, and Hesai XT-32. The sensors exhibit different sampling densities and spatial patterns while preserving largely overlapping scene structures, providing shared observations for cross-sensor representation learning.
Fig. 4: Representative scene types in SULID. Each column corresponds to a scene type, including structured, open, and dense environments. The top row shows accumulated point cloud maps (submaps), while the bottom row presents single-frame point clouds. Structured environments exhibit repetitive geometric patterns, open environments contain sparse structures with weak geometric constraints, and dense environments are characterized by strong occlusions and clutter. The red line indicates the driving trajectory. The red dot highlights the reference position.
Fig. 5: Range-dependent point distributions across different scene types. Point counts are computed within distance bins using Ouster OS1-128 scans and normalized by the nearest-range bin. The structured, open, and dense scenes exhibit distinct distribution patterns, reflecting differences in geometric sparsity, structural density, and occlusion characteristics.
Fig. 6: Training pipeline of LightLoc++. (a) In sensor-robust backbone learning, synchronized scans are transformed into a unified LiDAR coordinate system and fed into a shared backbone with multi-scene regression heads. Local and global consistency losses are imposed across synchronized LiDAR observations to learn sensor-robust representations. (b) In new-scene learning, the pretrained backbone is frozen, while lightweight scene-specific heads are trained with SCG and RSD for efficient SCR. LPC and WPC denote local point coordinates in the unified LiDAR coordinate system and world point coordinates, respectively. SC denotes sample classification.
Fig. 7: Network architectures used in LightLoc++. LightLoc++ uses the same backbone and prediction-head architectures as LightLoc [ 19 ] . Cls. head and Reg. head denote the sample classification head and the scene coordinate regression head, respectively. The backbone extracts scene-agnostic point features, the classification head provides spatial guidance for SCG, and the regression head predicts scene coordinates for pose estimation. Here, denotes the number of spatial clusters used for sample classification in SCG.
| Methods | Reference | Training Time | 15-13-06-37 | 17-13-26-39 | 17-14-03-00 | 18-14-14-42 | Average | |
| PointLoc [ 54 ] | Sens. J.’22 | 125 hrs. | 10.75 / 2.36 | 11.07 / 2.21 | 11.53 / 1.92 | 9.82 / 2.07 | 10.79 / 2.14 | |
| PosePN [ 20 ] | PR’22 | 5 hrs. | 9.47 / 2.80 | 12.98 / 2.35 | 8.64 / 2.19 | 6.26 / 1.64 | 9.34 / 2.25 | |
| PosePN++ [ 20 ] | PR’22 | 11 hrs. | 4.54 / 1.83 | 6.44 / 1.78 | 4.89 / 1.55 | 4.64 / 1.61 | 5.13 / 1.69 | |
| STCLoc [ 56 ] | TITS’23 | 20 hrs. | 5.14 / 1.27 | 6.12 / 1.21 | 5.32 / 1.08 | 4.76 / 1.19 | 5.34 / 1.19 | |
| NIDALoc [ 57 ] | TITS’24 | 20 hrs. | 3.71 / 1.50 | 5.40 / 1.40 | 3.94 / 1.30 | 4.08 / 1.30 | 4.28 / 1.38 | |
| HypLiLoc [ 55 ] | CVPR’23 | 17 hrs. | 5.03 / 1.46 | 4.31 / 1.43 | 3.61 / 1.11 | 2.61 / 1.09 | 3.89 / 1.27 |
TABLE III: Quantitative results on the QEOxford dataset. Mean ATE (m) / ARE ( ∘ ) are reported. . Best results in bold , second best results underlined . Training time is also included.
| Methods | Reference | Training Time | 2020-12-01 | 2021-02-02 | 2021-03-02 | 2021-04-29 | Average | |
| PointLoc [ 54 ] | Sens. J.’22 | 26 hrs. | 14.45 / 1.55 | 12.73 / 1.59 | 12.00 / 1.40 | 13.00 / 1.48 | 13.05 / 1.57 | |
| PosePN [ 20 ] | PR’22 | 6 hrs. | 15.59 / 4.18 | 5.58 / 0.99 | 4.59 / 0.86 | 5.82 / 0.88 | 7.90 / 1.73 | |
| PosePN++ [ 20 ] | PR’22 | 8 hrs. | 7.69 / 1.45 | 4.87 / 0.98 | 4.53 / 0.89 | 5.14 / 0.87 | 5.56 / 1.01 | |
| STCLoc [ 56 ] | TITS’23 | 13 hrs. | 4.64 / 0.73 | 4.56 / 0.80 | 4.38 / 0.80 | 4.71 / 0.63 | 4.57 / 0.74 | |
| NIDALoc [ 57 ] | TITS’24 | 21 hrs. | 5.01 / 0.87 | 4.87 / 0.82 | 4.87 / 0.82 | 4.78 / 0.71 | 4.88 / 0.81 | |
| HypLiLoc [ 55 ] | CVPR’23 | 8 hrs. | 4.46 / 0.70 | 1.86 / 0.71 | 1.66 / 0.53 | 1.57 / 0.44 | 2.39 / 0.60 |
TABLE V: Quantitative results on the Boreas dataset. Mean ATE (m) / ARE ( ∘ ) are reported. . Best results in bold , second best results underlined . Training time is also included.
| Methods | Reference | Training Time | DCC | |
| PointLoc [ 54 ] | Sens. J.’22 | 725 mins. | 16.38 / 5.16 | |
| PosePN [ 20 ] | PR’22 | 70 mins. | 18.30 / 7.68 | |
| PosePN++ [ 20 ] | PR’22 | 200 mins. | 6.64 / 3.43 | |
| STCLoc [ 56 ] | TITS’23 | 230 mins. | 7.89 / 3.25 | |
| NIDALoc [ 57 ] | TITS’24 | 321 mins. | 5.87 / 3.39 | |
| HypLiLoc [ 55 ] | CVPR’23 | 110 mins. | 8.96 / 2.64 |
TABLE VI: Quantitative results on the MulRan dataset. Mean ATE (m) / ARE ( ∘ ) are reported. . Best results in bold , second best results underlined . Training time is also included.
| Methods | Training Time | Library | Sports | |
| PointLoc [ 54 ] | 288 mins. | 6.25 / 3.15 | 4.45 / 4.62 | |
| PosePN [ 20 ] | 28 mins. | 7.54 / 4.65 | 4.23 / 5.26 | |
| PosePN++ [ 20 ] | 70 mins. | 4.70 / 2.75 | 5.82 / 4.93 | |
| STCLoc [ 56 ] | 98 mins. | 8.40 / 5.09 | 5.92 / 4.87 | |
| NIDALoc [ 57 ] | 113 mins. | 3.15 / 2.08 | 2.45 / 3.66 | |
| HypLiLoc [ 55 ] | 39 mins. | 5.81 / 2.07 | 4.87 / 3.65 |
TABLE VII: Quantitative results on the HeRCULES dataset. Mean ATE (m) / ARE ( ∘ ) are reported. The best and second-best results are highlighted in bold and underline , respectively. Average training time across two scenes is also included.
Fig. 8: Visualization results of predicted trajectories on QEOxford, Boreas, MulRan, and HeRCULES. The ground truth and prediction are black and red lines, respectively. The star denotes the first frame. The caption of each subfigure shows the mean ATE (m) and ARE ( ∘ ). For each trajectory, we highlight the best and second-best results.
| Pretraining | QEOxford | Boreas | MulRan | HeRCULES |
| KITTI | 1.50 / 1.73 | 2.75 / 1.17 | 5.33 / 4.32 | 2.09 / 2.98 |
| nuScenes | 0.83 / 1.12 | 2.12 / 1.24 | 4.76 / 2.77 | 2.27 / 2.92 |
| HeliPR | 0.95 / 1.27 | 1.31 / 0.82 | 3.62 / 2.50 | 2.13 / 2.95 |
| SULID | 0.72 / 1.03 | 0.98 / 0.66 | 3.50 / 2.39 | 2.02 / 2.81 |
| + | 0.69 / 1.02 | 0.92 / 0.66 | 3.56 / 2.41 | 1.83 / 2.62 |
| + | 0.70 / 0.99 | 0.87 / 0.64 | 3.36 / 2.18 | 1.75 / 2.60 |
TABLE VIII: Effect of multi-LiDAR representation learning. We evaluate localization performance across multiple LiDAR benchmarks using encoders pretrained under different settings. Results are reported as mean ATE (m) / ARE ( ∘ ).
| Method | O O | O A | O V | Average |
| Reference Localization Methods | ||||
| LiSA | 0.88 / 0.70 | 2.18 / 2.62 | 101.94 / 47.56 | 41.39 / 14.35 |
| RALoc | 1.95 / 1.01 | 22.05 / 26.01 | 246.13 / 120.23 | 90.04 / 49.08 |
| FlashMix | 6.76 / 5.69 | 245.61 / 73.77 | 164.89 / 100.68 | 104.32 / 60.05 |
| Our Framework with Different Pretrained Backbones | ||||
| KITTI | 1.74 / 1.14 | 38.22 / 25.63 | 213.73 / 103.25 | 84.56 / 43.34 |
TABLE IX: Cross-Sensor Generalization. LiSA, RALoc, and FlashMix are included as reference localization methods to indicate the cross-sensor generalization capability of existing approaches. For our framework, SCG and RSD are disabled, and only the pretrained encoder is varied. Results are reported as mean ATE (m) / ARE ( ∘ ).
Fig. 9: Cross-sensor localization generalization on the HeLiPR DCC sequence. Despite the substantial differences in sensing characteristics and fields of view, our method maintains accurate localization across Ouster, Aeva, and Velodyne sensors without fine-tuning. The star denotes the first frame.
| Backbone | SCG | RSD | Training Time | Mean Error [m/ ∘ ] | |
| 1 | 5.7 hrs. | 1.02 / 1.23 | |||
| 2 | 2.4 hrs. | 0.84 / 1.12 | |||
| 3 | 1.3 hrs. | 0.71 / 0.95 | |||
| 4 | 2 hrs. | 0.86 / 1.15 | |||
| 5 | 1 hrs. | 0.70 / 0.99 |
TABLE X: Ablation study . Training time required to achieve comparable accuracy on QEOxford. Backbone : multi-LiDAR backbone. SCG : sample classification guidance for regression learning. RSD : redundant sample downsampling for filtering well-learned samples during training.
| Clusters | 5 | 15 | 25 | 50 | 100 |
| Acc. | 100 | 99.89 | 99.76 | 99.62 | 99.17 |
| Err. | 0.76/1.06 | 0.73/1.05 | 0.70/0.99 | 0.73/1.02 | 0.77/1.07 |
TABLE XI: Effect of cluster number in SCG. We report the classification accuracy (%) and localization error (ATE/ARE) on the QEOxford dataset.
| Sampling | Random | Uniform | RSD |
| 15 | 0.84 / 1.21 | 0.77 / 0.98 | 0.71 / 0.97 |
| 25 | 0.96 / 1.27 | 0.83 / 1.00 | 0.70 / 0.99 |
| 35 | 1.12 / 1.43 | 1.04 / 1.32 | 0.82 / 1.02 |
TABLE XII: Impact of different sampling strategies and ratios . We report the mean error [m/ ∘ ] on the QEOxford dataset.
| Sequence | Length | Tag | Training | Testing |
| 11-14-02-26 | 9.37km | Sunny | ||
| 14-12-05-52 | 9.22km | Overcast | ||
| 14-14-48-55 | 9.04km | Overcast | ||
| 18-15-20-12 | 9.04km | Overcast | ||
| 15-13-06-37 | 8.85km | Overcast | ||
| 17-13-26-39 | 9.02km | Sunny |
TABLE XIII: Dataset Descriptions on the Oxford and QEOxford dataset.
Fig. 10: Visualization of the K-Means-based spatial partitions in SULID. Different colors denote different scene partitions, and the number in parentheses indicates the number of synchronized groups.