Geo$^\textbf{2}$: Geometry-Guided Cross-view Geo-Localization and Image Synthesis

Yancheng Zhang; Xiaohan Zhang; Guangyu Sun; Zonglin Lyu; Safwan Wshah; Chen Chen

cs.CVMar 26, 2026

Geo $^\textbf{2}$ : Geometry-Guided Cross-view Geo-Localization and Image Synthesis

Authors: Yancheng Zhang, Xiaohan Zhang, Guangyu Sun, Zonglin Lyu, Safwan Wshah, Chen Chen

Organizations: Institute of Artificial Intelligence, University of Central Florida · University of Vermont

Abstract

Cross-view geo-spatial learning consists of two important tasks: Cross-View Geo-Localization (CVGL) and Cross-View Image Synthesis (CVIS), both of which rely on establishing geometric correspondences between ground and aerial views. Recent Geometric Foundation Models (GFMs) have demonstrated strong capabilities in extracting generalizable 3D geometric features from images, but their potential in cross-view geo-spatial tasks remains underexplored. In this work, we present Geo^2, a unified framework that leverages Geometric priors from GFMs (e.g., VGGT) to jointly perform geo-spatial tasks, CVGL and bidirectional CVIS. Despite the 3D reconstruction ability of GFMs, directly applying them to CVGL and CVIS remains challenging due to the large viewpoint gap between ground and aerial imagery. We propose GeoMap, which embeds ground and aerial features into a shared 3D-aware latent space, effectively reducing cross-view discrepancies for localization. This shared latent space naturally bridges cross-view image synthesis in both directions. To exploit this, we propose GeoFlow, a flow-matching model conditioned on geometry-aware latent embeddings. We further introduce a consistency loss to enforce latent alignment between the two synthesis directions, ensuring bidirectional coherence. Extensive experiments on standard benchmarks, including CVUSA, CVACT, and VIGOR, demonstrate that Geo^2 achieves state-of-the-art performance in both localization and synthesis, highlighting the effectiveness of 3D geometric priors for cross-view geo-spatial learning.

Figures & tables

Figure 1: Illustration of directly using VGGT on satellite (a) and ground (c) images, leading to incorrect reconstructed shown in (b).

Method	CVGL	CVIS	Bi-dir	Joint	Geometric Prior
GeoDTR+ [ 53 ]	✓	✗	✗	✗	GLE [ 53 ]
Sample4Geo [ 8 ]	✓	✗	✗	✗	✗
RGCIS [ 44 ]	✗	✓	✓	✗	SAIG [ 60 ]
Sat2Density [ 26 ]	✗	✓	✗	✗	Volume Density
CDE [ 33 ]	✓	✓	✗	✓	SAFA [ 30 ]
Ours	✓	✓	✓	✓	VGGT [ 36 ]

Table 1: Comparison of representative cross-view geo-spatial learning methods on supported tasks and geometric priors.

Figure 2: Overview of the Geo ${}^{\textbf{2}}$ framework. We first extract geometric features from ground and satellite images using VGGT. These dense features are then embedded into a shared geometry-aware latent space as detailed in Sec. 3.2 . The resulting embeddings, $f^{g}$ and $f^{s}$ , are used for both CVGL and CVIS. While Geo ${}^{\textbf{2}}$ supports bidirectional image synthesis, it only requires training in the ground-to-satellite (G2S) direction. As detailed in Sec. 3.3 , only ground images are needed as input during inference for G2S generation, and vice versa.

Figure 3: Illustration of VGGT reconstructions for (a) the ground view and (b) the satellite view, showing strong geometric alignment (e.g., buildings and overall layout). The ground view reconstruction is obtained from four perspective crops, illustrated in (c).

Figure 4: Overview of GeoMap pipeline. Ground and satellite images are individually processed via two separate branches.

Figure 5: Overview of our GeoFlow pipeline. The latent representation ( $f^{g}$ if ground-to-satellite synthesis or $f_{s}$ if satellite-to-ground synthesis) is input as condition $C$ .

Dataset	Approach	R@1	R@5	R@10	R@1%	Hit Rate
	SAFA † [ 30 ]	33.93	58.42	68.12	98.24	36.87
	TransGeo [ 58 ]	61.48	87.54	91.88	99.56	73.09
VIGOR	SAIG-D [ 60 ]	65.23	88.08	-	99.68	74.11
Same-	GeoDTR [ 54 ]	56.51	80.37	86.21	99.25	61.76
area	GeoDTR+ [ 53 ]	59.01	81.77	87.10	99.07	67.41
	Sample4Geo [ 8 ]	77.86	95.66	97.21	99.61	89.82

Table 2: Comparison of cross-view geo-localization performance on the VIGOR dataset under same-area and cross-area settings. We report recall rates (%) and hit rate (%) at different top-

K

retrieval thresholds. The best results are shown in bold and the second-best results are underlined .

Approach	CVUSA				CVACT Val				CVACT Test
Approach	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%
LPN [ 40 ]	85.79	95.38	96.98	99.41	79.99	90.63	92.56	-	-	-	-	-
SAFA † [ 30 ]	89.84	96.93	98.14	99.64	81.03	92.80	94.84	-	-	-	-	-
TransGeo [ 58 ]	94.08	98.36	99.04	99.77	84.95	94.14	95.78	98.37	-	-	-	-
GeoDTR [ 54 ]	93.76	98.47	99.22	99.85	85.43	94.81	96.11	98.26	62.96	87.35	90.70	98.61
GeoDTR † [ 54 ]	95.43	98.86	99.34	99.86	86.21	95.44	96.72	98.77	64.52	88.59	91.96	98.74

Table 3: Comparison of cross-view geo-localization performance on CVUSA and CVACT datasets in recall at top-K retrieves (R@K). The best results are shown in bold and the second-best results are underlined . † indicates Polar Transformation is applied.

Dataset	Approach	R@1	R@5	R@10	R@1%
	SAFA [ 30 ]	30.40	52.93	62.29	85.82
	TransGeo [ 58 ]	37.81	61.57	69.86	89.14
CVUSA	SAIG-D [ 60 ]	15.29	33.07	42.14	72.95
$\downarrow$	GeoDTR [ 54 ]	43.72	66.99	74.61	91.83
CVACT	GeoDTR+ [ 53 ]	60.16	79.97	84.67	94.48
	Sample4Geo [ 8 ]	56.62	77.79	87.02	94.69

Table 4: Comparison of cross-view geo-localization performance on cross-dataset benchmarks. CVUSA

\rightarrow

CVACT stands for training on CVUSA and testing on CVACT. CVACT

\rightarrow

CVUSA stands for training on CVACT and testing on CVUSA. The best results are shown in bold and the second-best results are underlined .

Approach	CVUSA				CVACT				VIGOR
Approach	FID ( $\downarrow$ )	LPIPS ( $\downarrow$ )	PSNR ( $\uparrow$ )	SSIM ( $\uparrow$ )	FID ( $\downarrow$ )	LPIPS ( $\downarrow$ )	PSNR ( $\uparrow$ )	SSIM ( $\uparrow$ )	FID ( $\downarrow$ )	LPIPS ( $\downarrow$ )	PSNR ( $\uparrow$ )	SSIM ( $\uparrow$ )
X-Seq [ 28 ]	161.16	0.706	11.97	0.084	190.12	0.661	12.41	0.042	-	-	-	-
Aerial Diff [ 14 ]	136.18	0.855	10.06	0.103	127.29	0.878	10.24	0.108	123.16	0.831	11.49	0.141
GPG2A [ 2 ]	58.80	0.691	12.13	0.135	63.50	0.690	11.98	0.116	70.19	0.695	11.81	0.159
ControlNet [ 50 ]	32.45	0.650	12.63	0.149	62.21	0.682	11.95	0.115	53.27	0.666	10.38	0.170
Skydiffusion [ 46 ]	29.18	0.635	14.58	0.168	36.48	0.645	12.85	0.118	45.29	0.661	11.69	0.186

Table 5: Comparison of Ground-to-Satellite image synthesis performance on CVUSA, CVACT, and VIGOR datasets. We report FID [ 9 ] , LPIPS [ 51 ] , PSNR and SSIM scores. The best results are shown in bold .

Approach	CVUSA				CVACT				VIGOR
Approach	FID ( $\downarrow$ )	LPIPS ( $\downarrow$ )	PSNR ( $\uparrow$ )	SSIM ( $\uparrow$ )	FID ( $\downarrow$ )	LPIPS ( $\downarrow$ )	PSNR ( $\uparrow$ )	SSIM ( $\uparrow$ )	FID ( $\downarrow$ )	LPIPS ( $\downarrow$ )	PSNR ( $\uparrow$ )	SSIM ( $\uparrow$ )
Sat2Density [ 26 ]	41.43	0.4163	14.66	0.358	47.09	0.3339	16.38	0.482	47.98	0.3488	11.21	0.229
ControlNet [ 50 ]	51.48	0.5158	10.91	0.148	49.41	0.4563	11.66	0.229	53.29	0.4051	10.59	0.191
CrossViewDiff [ 5 ]	23.67	0.4412	12.00	0.371	41.94	0.3661	12.41	0.412	26.57	0.3414	13.68	0.268
Ours	29.05	0.4871	12.94	0.317	27.77	0.4833	13.57	0.457	22.90	0.4690	12.64	0.380

Table 6: Comparison of Satellite-to-Ground image synthesis performance on CVUSA, CVACT, and VIGOR datasets. We report FID [ 9 ] , LPIPS [ 51 ] , PSNR and SSIM scores. The best results are shown in bold .

Figure 6: Visualization of Generated images from Geo ${}^{\textbf{2}}$ on CVUSA dataset. From left to right are the ground truth satellite image, the ground truth ground image, the Ground-to-Satellite generated image, and the Satellite-to-Ground generated image.

Figure 7: Visualization of Generated images from Geo ${}^{\textbf{2}}$ on CVACT dataset. The order is the same as Figure 6 .

Figure 8: Visualization of Generated images from Geo ${}^{\textbf{2}}$ on VIGOR dataset. The order is the same as Figure 6 .

Dataset	#Ground Image	Ground Image Res.	#Satellite Image	Satellite Image Res.	Vertical FOV	Ground Crop Res.
CVUSA	44,416	1232 $\times$ 224	44,416	750 $\times$ 750	$\sim$ 90 ∘	224 $\times$ 224
CVACT	128,334	1664 $\times$ 832	128,334	1200 $\times$ 1200	180 ∘	416 $\times$ 416
VIGOR	105,214	2048 $\times$ 1024	90,618	640 $\times$ 640	180 ∘	512 $\times$ 512

Table 7: Panorama-to-satellite configuration across VIGOR, CVUSA, and CVACT datasets.

Dataset	Approach	R@1	R@5	R@10	R@1%
CVACT Val	Baseline	90.53	96.61	97.53	98.78
	+Geometry	93.81	97.02	97.77	98.92
	+ $\mathcal{L}_{KL}$	94.36	97.41	97.97	99.05
CVACT Test	Baseline	71.51	92.42	94.45	98.70
	+Geometry	74.37	93.24	95.02	98.86
	+ $\mathcal{L}_{KL}$	75.08	94.89	95.77	99.01

Table 8: Albaltion of CVGL task on CVACT dataset. “+Geometry” denotes model incorporating geometry features from VGGT are used. “

+\mathcal{L}_{KL}

” denotes model finetuned with consistency loss.

Direction	Approach	FID ( $\downarrow$ )	LPIPS ( $\downarrow$ )	PSNR ( $\uparrow$ )	SSIM ( $\uparrow$ )
G2S	Baseline	37.51	0.613	13.17	0.143
	+Geometry	33.02	0.568	14.08	0.155
	+ $\mathcal{L}_{KL}$	31.72	0.552	14.62	0.162
S2G	Baseline	33.41	0.573	11.08	0.377
	+Geometry	30.86	0.511	11.93	0.408
	+ $\mathcal{L}_{KL}$	27.77	0.483	13.57	0.457

Table 9: Albaltion of CVIS task on CVACT dataset. “+Geometry” denotes model incorporating geometry features from VGGT. “

+\mathcal{L}_{KL}

” denotes model finetuned with consistency loss. G2S is ground-to-satellite synthesis, and S2G is satellite-to-ground synthesis.

Figure 9: Illustration of the Equirectangular-to-Perspective transformation. Perspective-view crops are obtained by placing cameras with different yaw angles $\theta$ .

Dataset	Approach	R@1	R@5	R@10	R@1%	Hit Rate
Same-area	TransGeo	61.48	87.54	91.88	99.56	73.09
	Sample4Geo	77.86	95.66	97.21	99.61	89.82
	Ours $\dagger$	68.33	91.06	93.14	99.59	76.34
	Ours	81.59	96.53	98.62	99.68	90.35
Cross-area	TransGeo	18.99	38.24	46.91	88.94	21.21
	Sample4Geo	61.70	83.50	88.00	98.17	69.87

Table 10: Ablation on different backbones on VIGOR dataset. Ours

\dagger

denotes GeoMap with TransGeo as the semantic backbone, and ours denotes GeoMap with Sample4Geo as backbone.

Figure 10: Illustration of (a) accurate reconstruction from perspective crops, and (b) inaccurate reconstruction from directly cropped panoramas without E2P transformation.

Figure 11: Visualization of two satellite images generated from the same ground image from our Geo ${}^{\textbf{2}}$ and GPG2A [ 2 ] , respectively, on the VIGOR dataset. Although the results demonstrate similar PSNR and SSIM values, our generated satellite image is more visually and geometrically similar to the ground truth satellite image than GPG2A’s result. From left to right are the ground truth satellite image, GPG2A’s generated satellite image, our Geo ${}^{\textbf{2}}$ ’s generated satellite image, and the input ground image.

Figure 12: Visualization of the change of the PSNR and SSIM values with respect to the different levels of noise and vertical pixel shift.

	CVACT Val				CVACT Test
Approach	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%
Baseline	90.35	96.61	97.53	98.78	71.51	92.42	94.45	98.70
Baseline + Augmentation	91.56	96.94	97.71	98.83	72.83	93.15	94.74	98.81

Table 11: Evaluation of augmenting Cross-View Geo-Localization methods by using generated images from the proposed Geo

{}^{\textbf{2}}

on CVACT [ 19 ] Val and CVACT Test sets.

Figure 13: VGGT [ 36 ] reconstructions on VIGOR images under varying orientation angles and input orders.

$\theta$	R@1	R@5	R@1%
$\pi$ / $4$	94.36	97.41	99.05
$\pi$ / $8$	94.25	97.42	99.05
0	94.31	97.43	99.04

Table 12: Ablation study of different base navigation angle

\theta

on CVACT Val set.

Module	Sat.	Grd.	#Params
Conv Module	45	31	88M
VGGT	1387	5548	942M
GeoMap	40	158	43M

Table 13: Efficiency (in GFLOPS) and number of Parameters of our Geo

{}^{\textbf{2}}

Steps	FPS	GFLOPs	S2G FID/LPIP		G2S FID/LPIP
2	3.21	1143	29.09	0.4896	39.69	0.5594
5	1.38	2857	27.82	0.4831	33.65	0.5530
10	0.71	5715	27.77	0.4833	31.72	0.5520

Table 14: Ablation on ODE steps on CVACT Val set.

Figure 14: More Visualizations of Top-5 Retrieval Results from the CVACT Dataset. We compare the top-5 retrieval results of our method with the Sample4Geo [ 8 ] baseline. The query ground images are shown on the left, and the retrieval results are shown on the right.

Figure 15: More Visualization of Generated images from Geo ${}^{\textbf{2}}$ on CVUSA dataset. From left to right are the ground truth satellite image, the ground truth ground image, the Ground-to-Satellite generated image, and the Satellite-to-Ground generated image.

Figure 16: More Visualization of Generated images from Geo ${}^{\textbf{2}}$ on CVACT dataset. From left to right are the ground truth satellite image, the ground truth ground image, the Ground-to-Satellite generated image, and the Satellite-to-Ground generated image.

Figure 17: More Visualization of Generated images from Geo ${}^{\textbf{2}}$ on VIGOR dataset. From left to right are the ground truth satellite image, the ground truth ground image, the Ground-to-Satellite generated image, and the Satellite-to-Ground generated image.

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

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
Beyond 2D Matching: A Unified Single-Stage Framework for Geometry-Aware Cross-View Object Geo-Localization
Jun 29, 2026Liyao Wang, Ruipu Wu, Haojun Xu +3Cross-View Geo-Localization Camera Pose Estimation
Beihang University, Beijing, China · Meituan, Beijing, China
GeoLink: A 3D-aware Framework to Improve Generalization for Cross-view Geo-localization
Apr 14, 2026Hongyang Zhang, Yinhao Liu, Haitao Zhang +4Cross-View Geo-Localization Geospatial Representation Learning
Xiamen University · Tongji University

Oct 8, 2026J/K move · Enter open · S save

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

Ziqian Mo, Hill Zhang, Haosheng Tan, Ling Li+1

The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China · Claremont McKenna College Claremont, California, USA

Cross-View Geo-Localization Geospatial Representation Learning

Open paper AskPDF arXiv

Cross-view geo-localization (CVGL) aims to estimate geographic locations by matching images captured from different viewpoints, such as drone and satellite views. Existing methods mainly rely on visual representations, but often fail to jointly model fine-grained structural correspondences and semantic priors, making them prone to confusion between visually similar but semantically different regions, and thus limiting robustness under large viewpoint variations. To address these challenges, we propose \textbf{S

^3

Geo}, a structure-semantic synergistic learning framework for cross-view matching. Specifically, we first introduce a Decoupled Query Pooling (DQP) module to extract a compact set of region-aware features from dense tokens, enabling explicit modeling of local structural patterns. We then design a query-level contrastive learning scheme with an optimal transport (OT)-based formulation to establish soft correspondences under cross-view spatial misalignment. Furthermore, we incorporate a Semantic Knowledge Distillation (SKD) strategy from a frozen CLIP teacher to transfer semantic priors and relational structures, thereby improving discrimination on hard negatives. By operating synergistically, the semantic priors provide robust contextual filtering, which guides the structural module to establish precise spatial alignments. Experiments on the University-1652 and SUES-200 datasets demonstrate that \textbf{S

^3

Geo} consistently outperforms state-of-the-art approaches without increasing inference complexity, validating the effectiveness of jointly modeling structural and semantic information for CVGL.

Abstract

Cross-View Geo-Localization

Explore similar work

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

Beyond 2D Matching: A Unified Single-Stage Framework for Geometry-Aware Cross-View Object Geo-Localization

GeoLink: A 3D-aware Framework to Improve Generalization for Cross-view Geo-localization

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