SinGeo: Unlock Single Model's Potential for Robust Cross-View Geo-Localization

Yang Chen; Xieyuanli Chen; Junxiang Li; Jie Tang; Tao Wu

cs.CVMar 10, 2026

SinGeo: Unlock Single Model's Potential for Robust Cross-View Geo-Localization

Authors: Yang Chen, Xieyuanli Chen, Junxiang Li, Jie Tang, Tao Wu

Organizations: College of Intelligence Science and Technology, National University of Defense Technology

Abstract

Robust cross-view geo-localization (CVGL) remains challenging despite the surge in recent progress. Existing methods still rely on field-of-view (FoV)-specific training paradigms, where models are optimized under a fixed FoV but collapse when tested on unseen FoVs and unknown orientations. This limitation necessitates deploying multiple models to cover diverse variations. Although studies have explored dynamic FoV training by simply randomizing FoVs, they failed to achieve robustness across diverse conditions -- implicitly assuming all FoVs are equally difficult. To address this gap, we present SinGeo, a simple yet powerful framework that enables a single model to realize robust cross-view geo-localization without additional modules or explicit transformations. SinGeo employs a dual discriminative learning architecture that enhances intra-view discriminability within both ground and satellite branches, and is the first to introduce a curriculum learning strategy to achieve robust CVGL. Extensive evaluations on four benchmark datasets reveal that SinGeo sets state-of-the-art (SOTA) results under diverse conditions, and notably outperforms methods specifically trained for extreme FoVs. Beyond superior performance, SinGeo also exhibits cross-architecture transferability. Furthermore, we propose a consistency evaluation method to quantitatively assess model stability under varying views, providing an explainable perspective for understanding and advancing robustness in future CVGL research. Codes will be available upon acceptance.

Figures & tables

Figure 1: Comparison of different training paradigms for Cross-view Geo-localization. SinGeo achieves superior overall performance and better transferability under scenarios of both unknown orientation and different limited field of views. Previous methods including DSM [ 20 ] , SEH [ 5 ] , SAIG-D [ 34 ] , and ConGeo [ 13 ] are reported with their Top-1 Recall performance on CVUSA [ 29 ] dataset.

Figure 2: The proposed dual discriminative learning architecture. “Dual” denotes the two-branch design that enhances intra-view discriminativeness through self-supervision in both ground and satellite branches. Specifically, each branch applies specific transformations to generate $I_{g}^{*}$ and $I_{s}^{*}$ . The learning objective combines contrastive losses for intra-view discrimination and cross-view alignment.

Figure 3: Illustration of the curriculum learning framework and its inspiration. Left: An intuitive analogy of curriculum learning, where a freshman gradually refines geo-localization ability when he gets more familiar and practiced. Right:. The predetermined curriculum schedules the difficulties according to the epoch $t$ . After updated by dual discriminative learning, the encoder is then fed into Dynamic Similarity Sampling block [ 3 ] to generate negatives for next epoch.

Figure 4: Visualization of consistency comparison among SinGeo, ConGeo and Sample4Geo on the CVUSA dataset. The left column depicts a sample under varying orientations, where SinGeo exhibits stronger consistency in both ground and satellite branches. The right column presents a sample under different FoVs, where the activation heatmaps of SinGeo remain stable across varying views.

Set	Methods	Avg.	FoV $360^{\circ}$				FoV $180^{\circ}$				FoV $90^{\circ}$				FoV $70^{\circ}$
Set	Methods	R@1	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%
CVUSA	CVM-Net [ 6 ]	7.3	16.3	38.9	49.4	88.1	7.4	22.5	32.6	75.4	2.8	10.1	16.7	55.5	2.6	9.3	15.1	21.8
	CVFT [ 21 ]	10.0	23.4	44.4	55.2	86.6	8.1	24.3	34.5	75.2	4.8	14.8	23.2	61.2	3.8	12.4	19.3	55.6
	SEH [ 5 ]	40.9	85.4	93.5	95.8	-	53.7	72.3	79.0	-	16.6	32.2	40.3	-	7.8	18.8	25.6	-
	SAIG-D [ 34 ]	43.0	72.0	90.2	94.0	99.1	52.5	78.1	85.8	97.7	26.7	50.2	59.8	86.6	20.9	41.4	51.2	80.4
	GAL [ 16 ]	-	-	-	-	-	48.9	69.9	78.5	95.7	22.5	44.4	54.2	83.6	15.2	32.9	42.1	75.2

Table 1: Comparison with single-model performance of state-of-the-art methods on CVUSA and CVACT Val datasets under unknown orientation and limited FoV settings. The results of methods with multiple FoV-specialized trained models are reported as reference with gray background . Bold indicates the best performance among single-model methods . Underline denotes the best performance across all methods. “–” denotes metrics not provided in the original paper.

Methods	CVUSA		CVACT Val		CVACT Test
Methods	R@1	R@1%	R@1	R@1%	R@1	R@1%
DSM [ 20 ]	92.0	99.7	82.5	97.3	-	-
DeHi [ 25 ]	94.3	99.8	85.0	98.6	-	-
TransGeo [ 32 ]	94.1	99.8	85.0	98.4	-	-
GeoDTR [ 30 ]	95.4	99.9	86.2	98.8	64.5	98.7
SAIG-D [ 34 ]	96.3	99.9	89.1	98.9	67.5	96.8

Table 2: Comparison of state-of-the-art methods on CVUSA and CVACT datasets with north-aligned and center-aligned panorama-satellite pairs. The second-best performance is underlined.

Methods	Cross-Area				Same-Area
	FoV $=360^{\circ}$		FoV $=90^{\circ}$		FoV $=360^{\circ}$		FoV $=90^{\circ}$
	R@1	R@1%	R@1	R@1%	R@1	R@1%	R@1	R@1%
VIGOR [ 33 ]	1.4	44.6	-	-	19.1	95.1	-	-
TransGeo [ 32 ]	5.5	66.9	-	-	47.7	99.3	-	-
Sample4Geo [ 3 ]	9.0	43.7	0.5	21.6	14.2	54.9	1.1	30.6
ConGeo [ 13 ]	16.2	72.9	3.9	54.3	61.9	98.4	8.5	68.7

Table 3: Comparison of methods on cross-area and same-area splits of VIGOR under unknown orientation and Limited FoV settings.

Methods	St2S		S2St
Methods	R@1	R@10	R@1	R@10
University-1652 [ 31 ]	0.6	5.51	0.9	6.0
LPN [ 26 ]	0.7	-	1.4	-
ConGeo [ 13 ]	5.9	-	6.8	-
SinGeo	5.9	23.73	7.1	22.8

Table 4: Comparison of methods on St2S and S2St tasks of University-1652 dataset.

Figure 5: Qualitative evaluation on a sample of CVUSA dataset. Left: A query image under varied orientations and FoVs, together with its ground-truth satellite image. Right: Top-1 retrieved images and activated heatmap regions of different methods. Green circles denote the regions on the satellite image that correspond to the limited-FoV ground images.

Method	$\text{OC}_{\text{grd}}$	$\text{OC}_{\text{sat}}$	$\text{FC}_{\text{grd}}$	$\text{FC}_{\text{sat}}$
Sample4Geo [ 3 ]	0.65	0.67	0.50	0.66
ConGeo [ 13 ]	0.38	0.71	0.35	0.68
SinGeo	0.81	0.92	0.66	0.76

Table 5: Quantitative evaluation of orientation-consistency(OC) and FoV-consistency(FC) on CVUSA dataset.

Methods	FoV $=360^{\circ}$		FoV $=180^{\circ}$		FoV $=90^{\circ}$
Methods	R@1	R@1%	R@1	R@1%	R@1	R@1%
GeoDTR [ 30 ]	7.1	30.6	3.1	19.6	0.5	8.9
+ConGeo [ 13 ]	44.5	96.2	32.9	92.9	13.3	71.4
+SinGeo	54.8	98.2	37.1	95.9	17.8	82.8
Sample4Geo[CNN] [ 3 ]	93.3	99.1	84.6	99.5	55.1	96.6
+ConGeo [ 13 ]	85.2	98.9	92.3	99.7	55.9	90.9

Table 6: Cross-architecture generalization on the CVUSA dataset under unknown orientation and limited FoV settings.

DDL		CL	FoV $=360^{\circ}$		FoV $=180^{\circ}$		FoV $=90^{\circ}$
$I_{g}^{*}$	$I_{s}^{*}$	CL	R@1	R@1%	R@1	R@1%	R@1	R@1%
$\times$	$\times$	$\times$	93.3	99.1	84.6	99.5	55.1	96.6
$\checkmark$	$\times$	$\times$	85.2	98.9	92.3	99.7	55.9	90.9
$\checkmark$	$\checkmark$	$\times$	91.5	99.3	80.6	98.3	47.8	87.7
$\checkmark$	$\times$	$\checkmark$	96.2	99.5	92.1	99.5	66.9	96.4
$\checkmark$	$\checkmark$	$\checkmark$	96.8	99.5	91.8	99.5	70.1	97.5

Table 7: Ablations under different DDL and CL configurations on the CVUSA dataset.

Transformations	FoV $=360^{\circ}$		FoV $=180^{\circ}$		FoV $=90^{\circ}$
Transformations	R@1	R@1%	R@1	R@1%	R@1	R@1%
$T_{s}^{1}(\phi)$	89.2	97.3	77.5	94.2	63.7	96.4
$T_{s}^{2}(\phi)$	88.9	97.5	81.1	97.4	60.7	95.8
$T_{s}^{3}(p)$	96.8	99.5	91.8	99.5	70.1	97.5

Table 8: Performance comparison of different rotation transformations for

I_{s}

on the CVUSA dataset.

Scheduling Function	Avg.	FoV $=360^{\circ}$		FoV $=180^{\circ}$		FoV $=90^{\circ}$
Scheduling Function	R@1	R@1	R@1%	R@1	R@1%	R@1	R@1%
$f_{1}(x)$	86.2	96.8	99.5	91.8	99.5	70.1	97.5
$f_{2}(x)(\lambda=3)$	79.9	94.6	99.6	83.7	99.2	61.5	96.0
$f_{3}(x)(\lambda=3)$	82.8	97.1	99.8	90.9	99.3	60.4	92.3
$f_{2}(x)(\lambda=5)$	82.0	95.8	99.5	83.1	98.3	67.0	96.2
$f_{3}(x)(\lambda=5)$	82.3	96.6	99.8	89.8	99.6	60.5	94.5

Table 9: Performance comparison of different scheduling functions on the CVUSA dataset.

$\gamma$	$\omega_{1}\omega_{2}\omega_{3}$	Avg.	FoV $=360^{\circ}$		FoV $=180^{\circ}$		FoV $=90^{\circ}$
$\gamma$	$\omega_{1}\omega_{2}\omega_{3}$	R@1	R@1	R@1%	R@1	R@1%	R@1	R@1%
0.5	0.25	86.1	96.8	99.5	91.5	99.4	70.1	97.5
0.5	0.5	84.7	96.2	99.6	90.0	99.4	68.0	97.7
0.25	0.5	85.6	96.4	99.7	90.7	99.6	69.6	97.6
0.25	0.25	85.8	96.7	99.7	90.2	99.2	70.6	97.6

Table 10: Performance comparison of different loss weight configurations on the CVUSA dataset.

Methods	Avg.	FoV $360^{\circ}$				FoV $180^{\circ}$				FoV $90^{\circ}$				FoV $70^{\circ}$
Methods	R@1	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%
Sample4Geo [ 3 ]	2.4	8.0	12.5	14.0	31.0	1.0	2.7	3.8	25.1	0.4	1.3	2.2	24.1	0.2	0.7	1.1	16.9
ConGeo[360] [ 13 ]	25.0	64.3	81.5	84.7	95.6	28.6	47.8	54.9	88.4	5.2	13.0	17.9	67.8	1.8	5.6	8.4	53.3
ConGeo[180] [ 13 ]	25.4	37.4	55.3	61.1	89.6	45.6	67.1	73.0	94.3	13.7	28.3	35.6	81.7	5.0	13.1	18.0	68.6
SinGeo	35.8	66.7	84.2	86.8	96.0	46.0	66.8	72.6	93.2	19.6	38.8	47.1	89.3	11.0	25.5	33.3	84.2

Table 11: Comparison with single-model performance of selected methods on CVACT Test dataset under unknown orientation and limited FoV settings.

Set	Methods	FoV $360^{\circ}$				FoV $180^{\circ}$				FoV $90^{\circ}$				FoV $70^{\circ}$
Set	Methods	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%
Cross-Area	VIGOR [ 33 ]	1.4	-	-	44.6	-	-	-	-	-	-	-	-	-	-	-	-
	TransGeo [ 32 ]	5.5	-	-	66.9	-	-	-	-	-	-	-	-	-	-	-	-
	Sample4Geo [ 3 ]	9.0	-	-	43.7	-	-	-	-	0.5	-	-	21.6	-	-	-	-
	ConGeo [ 13 ]	16.2	-	-	72.9	-	-	-	-	3.9	-	-	54.3	-	-	-	-
	SinGeo	24.7	43.5	51.4	85.7	11.0	24.0	30.7	72.6	4.9	12.9	18.2	63.7	3.3	9.2	13.2	56.0

Table 12: Comparison of methods on cross-area and same-area splits of VIGOR under unknown orientation and limited FoV settings. “–” denotes metrics not provided in the original paper.

Metric	Method	$\text{OC}_{\text{grd}}$	$\text{OC}_{\text{sat}}$	$\text{FC}_{\text{grd}}$	$\text{FC}_{\text{sat}}$
Cosine Similarity	Sample4Geo [ 3 ]	0.37	0.40	0.25	0.32
	ConGeo [ 13 ]	0.15	0.61	0.22	0.46
	SinGeo	0.77	0.91	0.54	0.63
PCC	Sample4Geo [ 3 ]	0.13	0.18	0.04	0.11
	ConGeo [ 13 ]	0.10	0.46	0.17	0.21
	SinGeo	0.65	0.89	0.31	0.54

Table 13: Quantitative evaluation of orientation-consistency(OC) and FoV-consistency(FC) on CVUSA dataset based on cosine similarity and pearson correlation coefficient).

Figure 6: Qualitative visualization result on CVUSA. Yellow circles denote the regions on the satellite image that correspond to the limited-FoV ground images.

Figure 7: Qualitative visualization result on CVUSA. Yellow circles denote the regions on the satellite image that correspond to the limited-FoV ground images.

Figure 8: Qualitative visualization result on CVUSA. Yellow circles denote the regions on the satellite image that correspond to the limited-FoV ground images.

Cross-View Geo-Localization

Cross-view geo-localization matches ground, drone, or aerial images to maps or satellite views to estimate where they were captured.

See all 94 papers

Cross-view geo-localization, Image retrieval, Multiscale geometric modeling, Frequency domain enhancement
Mar 3, 2026Hongying Zhang, ShuaiShuai MaCross-View Image Matching Cross-View Geo-Localization
School of Electronic Information and Automation, Civil Aviation University of China, Tianjin, 300300, China
S $^3$ Geo: Structure-Semantic Synergistic Learning for Cross-View Geo-Localization
Oct 8, 2026Ziqian Mo, Hill Zhang, Haosheng Tan +2Cross-View Geo-Localization Geospatial Representation Learning
The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China · Claremont McKenna College Claremont, California, USA
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 Geospatial Representation Learning
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

Mar 3, 2026J/K move · Enter open · S save

Cross-view geo-localization, Image retrieval, Multiscale geometric modeling, Frequency domain enhancement

Hongying Zhang, ShuaiShuai Ma

School of Electronic Information and Automation, Civil Aviation University of China, Tianjin, 300300, China

Cross-View Image Matching Cross-View Geo-Localization Image Retrieval

Open paper AskPDF arXiv

Cross-view geo-localization (CVGL) aims to establish spatial correspondences between images captured from significantly different viewpoints and constitutes a fundamental technique for visual localization in GNSS-denied environments. Nevertheless, CVGL remains challenging due to severe geometric asymmetry, texture inconsistency across imaging domains, and the progressive degradation of discriminative local information. Existing methods predominantly rely on spatial domain feature alignment, which is inherently sensitive to large scale viewpoint variations and local disturbances. To alleviate these limitations, this paper proposes the Spatial and Frequency Domain Enhancement Network (SFDE), which leverages complementary representations from spatial and frequency domains. SFDE adopts a three branch parallel architecture to model global semantic context, local geometric structure, and statistical stability in the frequency domain, respectively, thereby characterizing consistency across domains from the perspectives of scene topology, multiscale structural patterns, and frequency invariance. The resulting complementary features are jointly optimized in a unified embedding space via progressive enhancement and coupled constraints, enabling the learning of cross-view representations with consistency across multiple granularities. Comprehensive experiments show that SFDE achieves competitive performance and in many cases even surpasses state-of-the-art methods, while maintaining a lightweight and computationally efficient design. {Our code is available at https://github.com/Mashuaishuai669/SFDE

Abstract

Cross-View Geo-Localization

Explore similar work

Cross-view geo-localization, Image retrieval, Multiscale geometric modeling, Frequency domain enhancement

S3^33Geo: Structure-Semantic Synergistic Learning for Cross-View Geo-Localization

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

S $^3$ Geo: Structure-Semantic Synergistic Learning for Cross-View Geo-Localization