Generalizable cross-view geo-localization aims to match the same location across views in unseen regions and conditions without GPS supervision. Its core difficulty lies in severe semantic inconsistency caused by viewpoint variation and poor generalization under domain shift. Existing methods mainly rely on 2D correspondence, but they are easily distracted by redundant shared information across views, leading to less transferable representations. To address this, we propose GeoLink, a 3D-aware semantic-consistent framework for Generalizable cross-view geo-localization. Specifically, we offline reconstruct scene point clouds from multi-view drone images using VGGT, providing stable structural priors. Based on these 3D anchors, we improve 2D representation learning in two complementary ways. A Geometric-aware Semantic Refinement module mitigates potentially redundant and view-biased dependencies in 2D features under 3D guidance. In addition, a Unified View Relation Distillation module transfers 3D structural relations to 2D features, improving cross-view alignment while preserving a 2D-only inference pipeline. Extensive experiments on multiple benchmarks show that GeoLink consistently outperforms state-of-the-art methods and achieves superior generalization across unseen domains and diverse weather environments.
Figures & tables
Figure 1. Performance comparisons on SUES-200@(150m). Existing methods perform well in the same-area setting but degrade noticeably in cross-area evaluation, revealing limited generalization ability under domain shift. Comparison plot showing same-area and cross-area performance for methods on SUES-200.
Figure 2. Comparison of previous methods and our approach. (a) Previous methods based only on 2D observations are vulnerable to occlusion, limited FoV, and ambiguous visual cues. (b) Scene-level 3D structure reconstructed from multi-view drone images provides a more stable structural anchor for learning transferable 2D retrieval features. Motivation illustration comparing view-biased two-dimensional cues with more stable three-dimensional structural cues.
Figure 3. The overview of the proposed framework. The key idea of GeoLink is to use scene-level 3D structure as a stable structural anchor for generalizable CVGL. Multi-view drone images are first converted into offline scene point clouds, which provide view-consistent structural priors during training. These 3D anchors guide 2D retrieval learning at two complementary levels: they regularize 2D features to reduce potentially view-biased dependencies, and they transfer instance-level structural relations from 3D space to 2D image features. Overview diagram of the GeoLink framework with 3D-guided linking, semantic refinement, and relation distillation.
Dataset
IDs
Images (D/S)
Altitude Range
University-1652
701
37,854 / 701
121m–256m
SUES-200
120
24,000 / 480
150m–300m
DenseUAV
2,256
6,768 / 13,536
80m–100m
Table 1. Statistics of the training CVGL benchmarks.
Method
Venue
U1652 → SUES
DUAV → SUES
SUES → U1652
DUAV → U1652
U1652 → DUAV
SUES → DUAV
R@1
AP
R@1
AP
R@1
AP
R@1
AP
R@1
AP
R@1
AP
MCCG
TCSVT’23
70.31
74.58
80.09
83.18
34.70
39.47
50.52
55.35
17.54
13.01
16.66
14.68
Sample4Geo
ICCV’23
82.03
85.12
78.94
81.86
62.13
66.48
37.90
42.54
33.46
22.28
24.56
17.63
DAC
TCSVT’24
87.65
89.80
86.44
88.70
61.52
65.62
45.41
50.03
37.81
26.99
25.37
18.02
MFRGN
ACM MM’24
78.28
81.75
67.22
71.14
35.58
40.95
37.95
41.73
23.60
16.08
22.55
14.51
CV-Cities
JSTARS’24
90.27
91.77
89.67
91.34
57.69
62.22
36.76
32.43
38.87
28.44
21.60
16.74
Table 2. Cross-area comparison (%) in the Drone → Satellite setting. Best and second-best results are highlighted in bold and underline , respectively.
Method
U1652 → SUES
DUAV → SUES
SUES → U1652
DUAV → U1652
R@1
AP
R@1
AP
R@1
AP
R@1
AP
MCCG
78.13
68.94
93.13
81.04
66.87
39.41
81.83
47.86
Sample4Geo
90.63
84.25
91.56
71.43
82.56
61.51
81.74
45.21
DAC
93.75
88.92
98.13
89.50
82.70
60.20
86.02
44.90
MFRGN
84.06
63.17
96.25
60.23
74.70
39.23
86.59
25.03
CV-Cities
96.56
88.07
98.41
88.13
78.31
56.85
83.58
71.66
Table 3. Cross-area comparison (%) in the Satellite → Drone setting. Best and second-best results are highlighted in bold and underline , respectively.
Method
Fog-Rain
Fog-Snow
Rain-Snow
Wind
R@1
AP
R@1
AP
R@1
AP
R@1
AP
Sample4Geo
37.40
44.13
26.68
33.05
45.81
51.58
59.35
64.58
DAC
73.23
76.72
45.03
50.57
68.97
73.07
86.23
88.39
CV-Cities
75.11
79.06
72.33
76.58
85.24
87.79
88.11
90.23
GeoLink
81.78
84.59
78.34
81.53
89.03
90.70
88.50
90.43
Table 4. Robustness evaluations on diverse weather cross-view transfer tasks (University-1652 → SUES-200, Drone → Satellite). Best and second-best results are highlighted in bold and underline , respectively.
Methods
SUES → DUAV
U1652 → SUES
R@1
AP
R@1
AP
Baseline ∗
15.56
11.89
62.78
60.19
Baseline ∗ +MME
17.86
13.95
73.55
71.21
Lga
Lsc
Lrd
R@1
AP
R@1
AP
✓
–
–
21.80
14.72
75.39
72.56
–
✓
–
26.68
19.25
92.97
88.62
Table 5. Ablation study of the major components in GeoLink under two settings: SUES → DUAV (Drone → Satellite) and U1652 → SUES (Satellite → Drone). ∗ denotes the method trained with the same parameters as GeoLink.
Figure 4. The impact of Lga on the InfoNCE losses between 3D and Drone views, 3D and Satellite views, respectively. Line charts of InfoNCE loss curves for 3D-drone and 3D-satellite view pairs.
Sample4Geo
DAC
CV-Cities
GeoLink
Params (M)
87.51
87.51
90.50
90.50
FLOPs (G)
72.60
80.22
91.67
91.72
Table 6. Comparisons of computational costs (Parameters and GFLOPs) of different methods.
Metric
University-1652
SUES-200
DenseUAV
BEV Similarity
0.69±0.07
0.78±0.05
0.66±0.06
VLM Judgement
99.29%
99.58%
97.78%
Table 7. Quality evaluation of VGGT-based Point Clouds.
Figure 5. The comparisons of PCA visualizations on robustness evaluation (University-1652 → SUES-200@150m). The discriminative objects are marked by rectangular boxes. PCA feature visualization comparing CV-Cities and GeoLink under adverse weather conditions.
Figure 6. The T-SNE comparisons on CV-Cities and GeoLink across eight locations (each color represents a location). Stars and crosses denote drone and satellite views, respectively. Blue circles represent false samples. t-SNE visualization of drone and satellite feature embeddings for eight locations.
Experts
SUES-200 → DenseUAV
SUES-200 → University
R@1
AP
R@1
AP
1
31.99
21.83
82.04
59.60
3
33.08
23.12
85.98
65.13
7
28.19
20.34
80.25
55.77
15
27.16
19.57
79.67
55.25
Table 8. Sensitivity Analysis of the number of experts in the MME block under two settings: SUES-200 → DenseUAV (Drone → Satellite), SUES-200 → University (Satellite → Drone). The best results are highlighted in bold.
Figure 7. Sensitivity analysis of (a) the trade-off hyper-parameter λ and (b) the hidden size of qθ(⋅) in vCLUB. Sensitivity plots for lambda and hidden size hyperparameters.