How to Reduce Localization Ambiguity? Geometry-Semantic Constrained BEV Representation Learning for Satellite-Ground Localization

• Insufficient radial-depth constraints leave ground features placeable at multiple distances along viewing rays, while similar appearances can make spatially distinct locations hard to distinguish.Source: §1 · p. 2, Figure 1 · p. 2
• GeoSem-BEV constrains BEV placement with radial-depth and vertical-height supervision, and descriptor learning with shared semantic supervision and same-class hard negatives.Source: §1 · p. 3
• The method is evaluated across three BEV localization frameworks and three datasets, with reduced orientation errors also reported on DReSS-D.Source: §1 · p. 3
• 37.2% and 38.1% lower mean orientation error on VIGOR unknown-orientation cross-area and same-area splits, respectively, with the lowest mean translation error in both.Source: §1 · p. 3

2.1 SATELLITE-GROUND LOCALIZATION

• Satellite-ground localization progressed from cross-view image retrieval to fine-grained pose estimation, with BEV matching becoming a major approach and correspondence-based methods making spatial associations explicit.Source: §2.1 · pp. 3–4

2.2 GEOMETRIC CONSTRAINTS FOR BEV REPRESENTATION CONSTRUCTION

• Existing BEV methods constrain vertical feature distribution but do not directly supervise radial depth, leaving features ambiguous along viewing rays.Source: §2.2 · p. 4

2.3 SEMANTIC INFORMATION FOR CROSS-VIEW MATCHING

• Semantic cues support cross-view association, but existing BEV descriptors rely mainly on appearance, matching, and pose supervision without explicit semantic supervision; same-class locations can still be spatially distinct.Source: §2.3 · pp. 4–5

3.1 OVERVIEW

• GeoSem-BEV adds geometry supervision during BEV construction and semantic supervision on descriptor fields, while each framework retains its correspondence estimator, pose solver, and base objective.Source: §3.1 · p. 5

3.2 GEOMETRY-CONSTRAINED BEV FEATURE CONSTRUCTION

• Radial-depth supervision weights sampled features according to predicted surface distance along a viewing ray, while vertical-height supervision constrains how samples contribute to each BEV cell.Source: §3.2 · p. 6
• Frozen monocular-depth teachers supply training targets for surface presence, radial distance, expected depth, uncertainty, and height; the image-based student predicts geometric weights at inference.Source: §3.2 · p. 6

3.3 EXPLICIT SEMANTIC SUPERVISION FOR BEV DESCRIPTORS

• SAM3-derived masks supervise a shared semantic head on ground and satellite descriptor fields, teaching descriptors to encode semantic content alongside appearance.Source: §3.3 · pp. 6–7
• Pose-aligned consistency and same-class hard negatives help distinguish spatially incorrect locations with similar semantics; the semantic head and targets are used only during training.Source: §3.3 · p. 7

3.4 CORRESPONDENCE SAMPLING AND TRAINING

• Each framework retains its correspondence estimator and base objective; weighted metric correspondences recover planar pose, while training adds geometric, semantic, and hard-negative losses.Source: §3.4 · p. 7

4.1 DATASETS AND EVALUATION METRICS

• The evaluation covers VIGOR same-area and cross-area splits, DReSS-D decentrality settings, and KITTI-CVL limited-field-of-view observations.Reported metrics include translation and orientation errors; VIGOR and KITTI-CVL also report localization or orientation recall.Source: §4.1 · p. 7

4.3 QUANTITATIVE RESULTS

• On VIGOR, all three GeoSem-BEV implementations reduce mean localization and orientation errors in both area splits; DReSS-D and KITTI-CVL also show improvements.GS-FG2 reduces cross-area mean translation error from 10.020 m to 4.662 m and mean orientation error from 31.410° to 13.850° relative to FG2.Source: Table 1 · p. 8, §4.3 · p. 8

4.2 IMPLEMENTATION DETAILS

• All variants retain their underlying frameworks’ correspondence and pose-estimation components while using GeoSem-BEV supervision.Configurations and adaptations are summarized in the appendices.Source: §4.2 · p. 8

4.4 ABLATION STUDY

• On VIGOR cross-area with unknown orientation, semantic-only supervision degrades all reported metrics, while geometry improves pose errors and the combined configuration achieves the highest recalls.Adding semantics and same-class hard negatives to geometric supervision further reduces mean translation and orientation errors.Source: Table 2 · p. 9, §4.4 · p. 9

4.5 DISCUSSION

• The magnitude and type of gains vary by framework and evaluation setting, and are clearest when the baseline retains substantial correspondence uncertainty.The discussion also reports that semantic supervision is most effective after geometric supervision has improved feature placement.Source: §4.5 · p. 9

5 CONCLUSION

• GeoSem-BEV combines radial-depth and vertical-height supervision with shared semantic supervision and same-class hard negatives to address two sources of localization ambiguity.On VIGOR with unknown orientation, it reduces mean orientation error by 37.2% cross-area and 38.1% same-area; on DReSS-D, reductions are 10.8% and 15.6%.Source: §5 · p. 9

ETHICS STATEMENT

• The work uses public cross-view localization benchmarks and reports no new personal-data collection or human-participant experiments.The authors say dataset licenses, image-access restrictions, and intended-use terms should be followed when distributing code or derived artifacts.Source: ETHICS STATEMENT · p. 10

AI USE STATEMENT

• The authors disclose AI assistance with writing, literature review, research execution, and manuscript drafting, while retaining responsibility for the research and final text.They state that AI-assisted literature suggestions and technical content are subject to author verification.Source: AI USE STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

• The experiments use Ascend NPUs, and Section 4 records evaluation protocols, metrics, and controlled training settings.Model adaptations are specified in Appendix B.6.Source: REPRODUCIBILITY STATEMENT · p. 10

APPENDIX

• The appendix is introduced as supplementary material, with a section on additional experiments.Source: APPENDIX · p. 14

A.2. KITTI-CVL Results.

• The appendix includes a section titled “A.3. Qualitative Results.”Source: §A.2 · p. 14
• A separate appendix section is titled “A.4. GT-Aligned BEV Correspondence Localization.”Source: §A.2 · p. 14
• The implementation appendix lists shared details, coordinate and sampling conventions, geometry and semantic supervision, training configuration, and pipeline adaptations.It names adaptations for GS-FG2, GS-ViewBridge, and GS-DenseFlow.Source: §A.2 · p. 14

• DReSS-D comparisons use test splits without RANSAC refinement, matching the published benchmark comparison.Source: §A.1 · p. 14, Table 3 · p. 14
• 10.8% and 15.6% lower mean orientation errors, and 7.7% and 11.3% lower localization errors, for GS-FG2 on DReSS-D unknown-orientation cross-area and same-area settings, respectively.Source: §A.1 · p. 14, Table 3 · p. 14
• GS-ViewBridge lowers unknown-orientation mean orientation error by 30.0% cross-area and 19.1% same-area, with slightly lower cross-area but higher same-area localization error.Source: §A.1 · p. 14

• In cross-area evaluation, GS-FG2 reduces both mean localization and orientation errors, while GS-ViewBridge reduces mean localization error.Source: §A.2 · p. 15
• In same-area evaluation, both implementations reduce mean orientation error and improve orientation recall, but their mean localization errors increase.Source: §A.2 · p. 15
• The results support GeoSem-BEV’s applicability to limited-field-of-view scenes, but improved orientation estimates do not guarantee lower localization error.Source: §A.2 · p. 15

• With radial-depth and vertical-height supervision, projected ground-view content is more spatially concentrated and better aligned with satellite structures than the unconstrained baseline.Source: §A.3 · p. 15, Figure 3 · p. 16
• Explicit semantic supervision concentrates matching responses around ground-truth locations and suppresses responses from visually similar regions elsewhere.Source: §A.3 · p. 15, Figure 3 · p. 16
• On challenging VIGOR samples, GS-FG2 produces more coherent correspondences and pose estimates closer to ground truth than FG2.Source: §A.3 · p. 15, Figure 4 · p. 16

A.4 GT-ALIGNED BEV CORRESPONDENCE LOCALIZATION

• On 5,000 VIGOR cross-area validation samples with unknown orientation, Geo-FG2 lowers mean correspondence error from 9.095 m to 8.427 m and median from 3.662 m to 3.244 m.It increases Corr.@1m, Corr.@3m, and Corr.@5m; a paired sample-level bootstrap gives a mean-error difference of −0.669 m (95% CI [−0.796, −0.539] m).Source: Table 5 · p. 17, §A.4 · p. 17

B.1 SHARED IMPLEMENTATION DETAILS

• The GS variants share radial geometric supervision, explicit semantic descriptor supervision, and same-class hard negatives, but differ in encoders, aggregation normalization, and correspondence estimators.Shared training and evaluation settings are summarized in Appendix B.5, with implementation-specific departures described separately.Source: §B.1 · p. 17

B.2 COORDINATE AND SAMPLING CONVENTIONS

• GS-FG2 uses a 41 × 41 horizontal grid spanning [−35.5, 35.5] m, 11 heights spanning [−10, 10] m, and consistent coordinate and transform conversions.Satellite offsets are converted to metres using ground sampling distance, and equirectangular sampling wraps horizontally.Source: §B.2 · p. 17

B.3 GEOMETRY SUPERVISION DETAILS

• Geometry targets combine teacher-derived metric pseudo-depth supervision with satellite-derived relative-height targets over query bins, whose height loss updates the surface and radial predictions.The ground-view teacher supplies surface, radial-distribution, expected-depth, and uncertainty targets; the height loss is normalized by log 11.Source: §B.3 · pp. 17–18

B.4 SEMANTIC SUPERVISION DETAILS

• Semantic supervision uses per-view soft targets and predicted class probabilities, then compares pose-warped ground predictions with satellite predictions through a consistency term.The target distributions are fixed supervision signals distinct from the student predictions.Source: §B.4 · p. 18
• The directional hard-negative loss selects the top 32 same-class candidates outside the one-cell neighbourhood for each anchor.It is applied in both ground-to-satellite and satellite-to-ground directions.Source: §B.4 · p. 18

B.5 TRAINING AND EVALUATION CONFIGURATION

• GS-FG2 experiments use frozen 1024-channel DINOv2 features, a 41 × 41 × 11 query grid, 64 radial bins, and 128-dimensional descriptors, with 25 epochs and global batch size 24 for the main VIGOR setting.The main experiment uses seed 0, unknown orientation, random rotations within [−180°, 180°], and the official 53,694-sample test split.Source: §B.5 · p. 18

B.6 PIPELINE ADAPTATIONS

• The evaluated adaptations are implementation-specific: GS-FG2 is described first, followed by GS-ViewBridge and GS-DenseFlow.The descriptions do not assert identical architectures across back ends.Source: §B.6 · p. 18

B.6.1 GS-FG2

• GS-FG2 retains FG2 descriptor projection, dual-softmax matching, and a correspondence-based pose solver, with satellite height targets supervising its likelihood weights.It samples frozen DINOv2 features on a metric 3D grid and aggregates appearance features through six planar self-attention/FFN blocks.Source: §B.6.1 · p. 19

B.6.2 GS-VIEWBRIDGE

• GS-ViewBridge replaces learned panorama cross-attention with analytic metric-grid sampling and combines radial likelihood with a separate vertical distribution for aggregation.Its ground mass also gates features and weights correspondence sampling and alignment; the model samples 1,024 pairs for Weighted Procrustes.Source: §B.6.2 · p. 19

B.6.3 GS-DENSEFLOW

• GS-DenseFlow forms geometry-constrained BEV features from ResNet18 features while retaining DenseFlow’s RAFT-based flow estimator.It normalizes appearance aggregation by column mass, retains unnormalized mass as geometric confidence, and applies semantic supervision to feature maps used to construct correlations.Source: §B.6.3 · p. 19