cs.CVMar 17, 2026
SaveBEV-SLD: Self-Supervised Scene Landmark Detection for Global Localization with LiDAR Bird's-Eye View Images
Organizations: Institute for Photogrammetry and Geoinformatics, University of Stuttgart, Germany
Abstract
We present BEV-SLD, a LiDAR global localization method building on the Scene Landmark Detection (SLD) concept. Unlike scene-agnostic pipelines, our self-supervised approach leverages bird's-eye-view (BEV) images to discover scene-specific patterns at a prescribed spatial density and treat them as landmarks. A consistency loss aligns learnable global landmark coordinates with per-frame heatmaps, yielding consistent landmark detections across the scene. Across campus, industrial, and forest environments, BEV-SLD delivers robust localization and achieves strong performance compared to state-of-the-art methods.
Figures & tables
Figure 2: Overview of the proposed landmark learning framework (left) and the localization pipeline at inference time (right).
Figure 3: Initial landmarks and landmarks after joint learning of landmark positions and detection in a factory floor environment: filled orange circles represent landmark positions at initialization and green non-filled circles after training. Orange lines connect corresponding landmarks. The background point cloud is shown for visualization only.
Figure 4: Network architecture. Res. Blocks represent residual blocks [ 12 ] , Down Blocks consist of max pooling followed by a residual block and Conv. Blocks consist of convolution followed by layer normalization and activation function.
Figure 5: Qualitative comparison of BEVPlace++ [ 23 ] trained on KITTI (K), LightLoc [ 19 ] , KISS-Matcher [ 22 ] , PosePN++ [ 38 ] and the proposed method. Top left : reference trajectory (ntu_day_01, orange) and ground truth trajectory of the test sequence (ntu_day_10, green). Others : inliers (errors 2 m, 5 ∘ ) of the different methods. Our method achieves broader inlier pose coverage than the baselines.
| Name | MCD [ 25 ] | NCLT [ 2 ] | Wild-Places [ 16 ] | Factory Floor |
| # Seq. | 6 | 8 | 4 | 2 |
| Acc. Dist. | 10.7 km | 43.3 km | 12.7 km | 875 m |
| Time Span | 13 days | 4 months | 14 months | 9 months |
| Area [m m] | ||||
| Environment | Campus | Campus | Forest | Industrial |
Table 1: Overview of datasets. Reported are # Seq., accumulated distance (Acc. Dist.), acquisitions time span, bounding box area, and environment type.
| Dataset | Seq. | # Frames | # Epochs | Duration [h] |
| MCD | ntu_day_01 | 5,694 | 600 | 14 |
| NCLT | 2012-01-15 | 15,026 | 300 | 12 |
| Wild-Places | V-03 | 6,757 | 800 | 16.5 |
| Factory Floor | 2024-01 | 1,031 | 50 | 1 |
Table 2: Training time per reference sequence.
| Seq. ntu_x (reference: day_01) | Frozen Initial Landmarks | Optimized Landmarks | ||||
| SR ↑ [%] | TE ↓ [m] | RE ↓ [deg] | SR ↑ [%] | TE ↓ [m] | RE ↓ [deg] | |
| day_02 | 68.05 | 1.09 | 1.63 | 100.00 | 0.19 | 0.27 |
| day_10 | 40.35 | 2.43 | 4.00 | 92.04 | 0.20 | 0.44 |
| night_04 | 38.65 | 2.59 | 4.06 | 91.82 | 0.21 | 0.44 |
| night_08 | 48.39 | 1.88 | 2.93 | 95.26 | 0.21 | 0.47 |
| night_13 | 53.64 | 1.64 | 2.39 | 99.91 | 0.20 | 0.39 |
Table 3: Localization results using frozen initial and optimized landmarks, evaluated on MCD dataset scenes.
Figure 6: Heatmap prediction for an unseen scene. Left : MCD dataset BEV image. Mid : Heatmap prediction using a Factory Floor dataset trained network. Right : Detected local peaks.
Figure 7: Success rate as a function of the distance to the closest reference pose, aggregated over all sequences. Distances are grouped into 3 m bins.
| Dataset (reference sequence) | Seq. | BEVPlace++ (K) [ 23 ] | BEVPlace++ (S) [ 23 ] | LightLoc [ 19 ] | KISS-Matcher [ 22 ] | PosePN++ [ 38 ] | BEV-SLD (Ours) | ||||||||||||
| SR ↑ [%] | TE ↓ [m] | RE ↓ [deg] | SR ↑ [%] | TE ↓ [m] | RE ↓ [deg] | SR ↑ [%] | TE ↓ [m] | RE ↓ [deg] | SR ↑ [%] | TE ↓ [m] | RE ↓ [deg] | SR ↑ [%] | TE ↓ [m] | RE ↓ [deg] | SR ↑ [%] | TE ↓ [m] | RE ↓ [deg] | ||
| MCD ntu_x | day_02 | 100.00 | 0.27 | 0.45 | 99.96 | 0.29 | 0.50 | 43.50 | 1.36 | 4.57 | 40.93 | 1.51 | 4.61 | 15.80 | 4.00 | 3.99 | 100.00 | 0.19 | 0.27 |
| (day_01) | day_10 | 81.53 | 0.22 | 0.54 | 80.24 | 0.26 | 0.59 | 42.72 | 1.20 | 5.41 | 43.71 | 1.25 | 5.01 | 13.35 | 4.54 | 3.73 | 92.04 | 0.20 | 0.44 |
| night_04 | 86.74 | 0.46 | 0.60 | 86.13 | 0.48 | 0.62 | 14.27 | 14.68 | 6.23 | 14.78 | 14.53 | 6.29 | 9.71 | 29.21 | 12.94 | 91.82 | 0.21 | 0.44 | |
| night_08 | 87.51 | 0.23 | 0.56 | 86.82 | 0.26 | 0.60 | 40.54 | 1.53 | 4.69 | 40.06 | 1.60 | 4.79 | 14.09 | 4.18 | 3.63 | 95.26 | 0.21 | 0.47 | |
| night_13 | 99.87 | 0.20 | 0.50 | 99.31 | 0.22 | 0.54 | 48.55 | 1.07 | 4.01 | 49.63 | 1.18 | 3.92 | 18.23 | 3.45 | 2.74 | 99.91 | 0.20 | 0.39 | |
Table 4: Quantitative results across datasets. We report median translational error (TE), median rotational error (RE), and success rate SR (errors 2 m, 5 ∘ ). Best results are highlighted in bold .
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Landmark initialization visualized for three exemplary keyframe poses.
Figure 9: Landmarks in green after self-supervised joint landmark detection and positioning on sequence 2012-01-15 of the NCLT dataset [ 2 ] in a campus environment. In total, 1592 landmarks are utilized in this scene. A point cloud of the entire scene in the background is used for illustration purposes.
Figure 10: Landmarks in green after self-supervised joint landmark detection and positioning on sequence V-03 of the Wild-Places dataset [ 16 ] in a forest environment. In total, 1757 landmarks are utilized in this scene. A point cloud of the entire scene in the background is used for illustration purposes.