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

Junming Feng; Panwang Xia; Qiong Wu; Xudong Lu; Zeyu Jiao; Kun Lv; Zherong Wu; Yi Wan

cs.CVSep 30, 2026

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

Authors: Junming Feng, Panwang Xia, Qiong Wu, Xudong Lu, Zeyu Jiao, Kun Lv, Zherong Wu, Yi Wan, +3 more

Organizations: The Hong Kong Polytechnic University, Hong Kong · Southern University of Science and Technology · Wuhan University, Wuhan, China · The Chinese University of Hong Kong, Hong Kong, China · Huawei Technologies Co., Ltd

Abstract

Satellite-ground localization estimates the planar position and yaw orientation of a ground camera within a geo-referenced satellite image. Most recent methods map ground and satellite features into a shared bird's-eye-view (BEV) space and establish spatial correspondences. However, insufficient depth constraints can assign one ground feature to different distances along a viewing direction, creating geometric ambiguity in BEV feature placement. Similar appearances at different locations can also create descriptor matching ambiguity, while existing descriptor learning lacks explicit semantic supervision to distinguish them. We propose GeoSem-BEV, a geometry-semantic constrained BEV representation learning method. Radial depth supervision constrains distance assignment, and vertical height supervision constrains height aggregation. Shared explicit semantic supervision promotes consistent semantic predictions across views and helps distinguish locations with similar semantics. These constraints improve feature placement and descriptor discriminability, enhancing state-of-the-art BEV localization models. On VIGOR with unknown orientation, GeoSem-BEV reduces mean orientation error by 37.2% and 38.1% in the cross-area and same-area settings, respectively. The corresponding errors are reduced by 10.8% and 15.6% on DReSS-D. On KITTI-CVL, it reduces same-area mean orientation error by 26.8% under 10 degree orientation noise.

Figures & tables

Figure 1: Two ambiguities in BEV-space feature matching. Left: Insufficient radial-depth constraints leave BEV feature placement ambiguous along viewing rays. (a) Upper right: Similar appearances at different locations can produce ambiguous descriptors without explicit semantic supervision. (b) Lower right: Explicit semantic supervision helps distinguish them.

Figure 2: GeoSem-BEV with FG 2 as the base model. Geometric supervision constrains radial assignment and height aggregation of ground-view features. A shared semantic head supervises both descriptor fields, and same-class hard negatives penalize spatially incorrect matches.

	Cross-area				Same-area
	Localization (m) $\downarrow$		Orientation ( ∘ ) $\downarrow$		Localization (m) $\downarrow$		Orientation ( ∘ ) $\downarrow$
Method	Mean	Median	Mean	Median	Mean	Median	Mean	Median
SliceMatch	7.220	3.310	25.970	4.510	6.490	3.130	25.460	4.710
CCVPE	5.410	1.890	27.780	13.580	3.740	1.420	12.830	6.620
DenseFlow	7.670	3.670	17.630	2.940	4.970	1.900	11.200	1.590
GS-DenseFlow	5.647 (-26.38%)	2.344 (-36.13%)	15.466 (-12.27%)	2.157 (-26.63%)	4.677 (-5.90%)	1.906 (+0.32%)	9.340 (-16.61%)	1.589 (-0.06%)

Table 1: VIGOR unknown-orientation test results. Lower is better. GeoSem-BEV implementations are shown in black bold; best, second-best, and third-best distinct values within each area split are marked in red bold, blue bold, and black bold, respectively. Ties share the same rank. FG 2 reports original single-stage results; published baselines follow the cited comparisons. For each GeoSem-BEV implementation, percentages in parentheses after every metric report the relative change from its corresponding base model, computed as

(E_{\mathrm{GS}}-E_{\mathrm{base}})/E_{\mathrm{base}}\times 100\%

Method	Loc. mean $\downarrow$	Loc. median $\downarrow$	Ori. mean $\downarrow$	Ori. median $\downarrow$	R@1m/5 ∘ $\uparrow$	R@3m/10 ∘ $\uparrow$	R@5m/20 ∘ $\uparrow$
	(m)	(m)	( ∘ )	( ∘ )
FG 2†	5.751	2.943	18.908	1.337	13.07	50.11	67.04
Sem-FG 2	6.339	3.335	22.740	1.397	11.06	45.25	62.76
Geo-FG 2	5.014	2.469	15.980	1.240	15.57	57.75	74.26
GS-FG 2	4.662	2.364	13.850	1.241	16.53	59.81	76.72

Table 2: Ablation on VIGOR cross-area with unknown orientation. All results use RANSAC. Best values in each column are bold. FG 2† denotes our reproduced FG 2 baseline.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Cross-area				Same-area
Method	Localization (m) $\downarrow$		Orientation ( ∘ ) $\downarrow$		Localization (m) $\downarrow$		Orientation ( ∘ ) $\downarrow$
	Mean	Median	Mean	Median	Mean	Median	Mean	Median
CCVPE	6.050	2.230	37.390	10.270	3.010	1.020	14.440	7.940
FG 2†	6.422	3.814	16.466	2.411	5.508	2.850	14.124	2.129
GS-FG 2	5.926 (-7.72%)	3.396 (-10.96%)	14.696 (-10.75%)	2.234 (-7.34%)	4.886 (-11.29%)	2.475 (-13.16%)	11.921 (-15.60%)	1.939 (-8.92%)
ViewBridge	4.220	2.370	12.580	2.600	2.940	1.490	8.470	1.620

Appendix

Table 3: DReSS-D test results under unknown orientation. All results are reported without RANSAC refinement, following the published DReSS-D comparison. CCVPE and ViewBridge values follow that comparison, and FG 2† denotes our re-implementation. Best, second-best, and third-best distinct values within each metric are marked in red bold, blue bold, and black bold, respectively; ties share the same rank. Percentages in parentheses after every metric report the relative change from the corresponding base model, computed as

(E_{\mathrm{GS}}-E_{\mathrm{base}})/E_{\mathrm{base}}\times 100\%

Area	Ori.	Method	Loc. (m) $\downarrow$		Ori. ( ∘ ) $\downarrow$		Ori. (%) $\uparrow$
			Mean	Median	Mean	Median	R@1 ∘	R@5 ∘
Cross-area	$\pm 10^{\circ}$	GGCVT	–	–	–	–	98.98	100.00
		CCVPE	9.160	3.330	1.550	0.840	57.72	96.19
		HC-Net	8.470	4.570	3.220	1.630	33.58	83.78
		FG 2	7.310	4.150	3.620	2.370	23.03	77.84
		GS-FG 2	7.026 (-3.89%)	4.372 (+5.35%)	3.485 (-3.73%)	2.003 (-15.49%)	28.07 (+21.88%)	80.48 (+3.39%)

Appendix

Table 4: KITTI-CVL test results. Orientation noise is sampled uniformly within

\pm 10^{\circ}

during training and testing. Lateral and longitudinal recalls are omitted. Best, second-best, and third-best distinct values within each area and orientation setting are marked in red bold, blue bold, and black bold, respectively; ties share the same rank. For GS-FG 2 and GS-ViewBridge, percentages in parentheses after every metric report relative changes from the corresponding base model. A dash denotes an unreported source metric.

Figure 3: Qualitative visualization of geometry-constrained feature placement and explicitly supervised descriptor learning. In the left block, the orange overlay is the RGB content sampled from the selected ground-image region and projected onto the satellite images; the two columns of satellite images show the projection produced by FG 2† and Geo-FG 2 , respectively. The cyan $+$ and arrow mark the ground-truth camera position and orientation used for alignment. In the right block, each heatmap shows the normalized descriptor-matching response for one ground-view BEV query over satellite locations: warmer colors indicate higher matching scores. The green triangle marks the ground-truth satellite location corresponding to the query cell, and the white star marks the highest-scoring predicted location; the reported peak error is the distance between these two markers.

Figure 4: Qualitative correspondence comparison on challenging VIGOR samples. Each row shows the same ground–satellite image pair for FG 2 (left) and GS-FG 2 (right). Lines visualize selected correspondences; green and red arrows indicate the ground-truth and predicted camera poses, respectively.

Method	Valid-cell ratio	Corr. error mean $\downarrow$ (m)	Corr. error median $\downarrow$ (m)	Corr.@1m $\uparrow$	Corr.@3m $\uparrow$	Corr.@5m $\uparrow$
FG 2†	0.726	9.095	3.662	9.29	41.20	60.40
Geo-FG 2	0.726	8.427	3.244	10.64	45.52	64.86

Appendix

Table 5: Ground-truth-aligned BEV correspondence localization on 5,000 VIGOR cross-area validation samples with unknown orientation. Correspondence errors are measured from the highest-scoring satellite-view BEV cell to the aligned location. Recalls are percentages. Best values are in bold.

Explore similar work

BGG: Bridging the Geometric Gap between Cross-View images by Vision Foundation Model Adaptation for Geo-Localization
May 11, 2026Wei Wang, Dou Quan, Ning Huyan +4Cross-View Geo-Localization Cross-View
Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education of China, Xidian University · Department of Automation, Tsinghua University, Beijing 100084, China · School of Telecommunications, Xidian University, Xi’an 710071, China
More Than Where You Are: Learning Semantics, Structure, and Geometry from Cross-View Localization
Jul 14, 2026Mao Chen, Xiangkai Zhang, Zhiyong Liu +2Cross-View Sequential Visual Localization
State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, and also with the School of Artificial Intelligence, University of Chinese Academy of Sciences · Beijing Aerospace Control Center
Warp-free Cross-view Geo-localization via Feature-space Consensus Mining
Aug 10, 2026Zhuo Song, Lian Xu, Runqing Jiang +4Cross-View Geo-Localization Cross-View
Sun Yat-sen University, Shenzhen, China · The University of Western Australia, Perth, Australia

May 11, 2026, cs.CVJ/K move · Enter open · S save

BGG: Bridging the Geometric Gap between Cross-View images by Vision Foundation Model Adaptation for Geo-Localization

Wei Wang, Dou Quan, Ning Huyan, Shuang Wang+3

Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education of China, Xidian University · Department of Automation, Tsinghua University, Beijing 100084, China · School of Telecommunications, Xidian University, Xi’an 710071, China

Cross-View Geo-Localization Cross-View Feature Extraction Frequency-Aware Regularization Scale Gap

Open paper AskPDF arXiv

Geometric differences between cross-view images, such as drone and satellite views, significantly increase the challenge of Cross-View Geo-Localization (CVGL), which aims to acquire the geolocation of images by image retrieval. To further enhance the CVGL performance, this paper proposes a parameter-efficient adaptation framework for bridging the geometric gap across images based on the vision foundation model (VFM) (e.g., DINOv3), termed BGG. BGG not only effectively leverages the general visual representations of VFM and captures the robust and consistent features from cross-view images, but also utilizes the generalization capabilities of the VFM, significantly improving the CVGL performance. It mainly contains a Multi-granularity Feature Enhancement Adapter (MFEA) and a Frequency-Aware Structural Aggregation (FASA) module. Specifically, MFEA enhances the scale adaptability and viewpoint robustness of features by multi-level dilated convolutions, effectively bridging the cross-view geometric gap with small training costs. Additionally, considering the [CLS] token lacks spatial details for precise image retrieval and localization, the FASA module modulates patch tokens in the frequency domain and performs adaptive aggregation for local structural feature enhancement. Finally, BGG fuses the enhanced local features with the [CLS] token for more accurate CVGL. Extensive experiments on University-1652 and SUES-200 datasets demonstrate that BGG has significant advantages over other methods and achieves state-of-the-art localization performance with low training costs.

Abstract

Explore similar work

BGG: Bridging the Geometric Gap between Cross-View images by Vision Foundation Model Adaptation for Geo-Localization

More Than Where You Are: Learning Semantics, Structure, and Geometry from Cross-View Localization

Warp-free Cross-view Geo-localization via Feature-space Consensus Mining