Organizations: College of Computer Science and Technology, Zhejiang University · State Key Laboratory of Blockchain and Data Security, Zhejiang University · Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security
Recent advances in cross-view geo-localization (CVGL) methods have shown strong potential for supporting unmanned aerial vehicle (UAV) navigation in GNSS-denied environments. However, existing work predominantly focuses on matching UAV views to onboard map tiles, which introduces an inherent trade-off between accuracy and storage overhead, and overlooks the importance of the UAV's heading during navigation. Moreover, the substantial discrepancies and varying overlaps in cross-view scenarios have been insufficiently considered, limiting their generalization to real-world scenarios. In this paper, we present Bearing-UAV, a purely vision-driven cross-view navigation method that jointly predicts UAV absolute location and heading from neighboring features, enabling accurate, lightweight, and robust navigation in the wild. Our method leverages global and local structural features and explicitly encodes relative spatial relationships, making it robust to cross-view variations, misalignment, and feature-sparse conditions. We also present Bearing-UAV-90k, a multi-city benchmark for evaluating cross-view localization and navigation. Extensive experiments show encouraging results that Bearing-UAV yields lower localization error than previous matching/retrieval paradigm across diverse terrains. Our code and dataset will be made publicly available.
Figures & tables
Figure 1: Bearing-UAV overview. Given a UAV-view patch (UVP) and four adjacent remote-sensing tile (RST) features with relative coordinates ( C ), the model jointly regresses position and heading.
Figure 2: Bearing-UAV in training mode. Given four adjacent RSTs with their relative coordinates ( C ) and one UVP within the same remote-sensing block (RSB), Bearing-UAV predicts the UAV’s absolute position and heading angle. Two GLUF submodules with shared parameters extract UVP/RST features. RCE encodes the relative coordinate of each RST. CA captures overlap-aware cross-view correspondences via cross-attention between UVP and RST features. PSG estimates a weighted position vector by aggregating UVP-RSTs patch similarities. We then fuse these output features via a residual-style connection, and apply two independent heads to jointly regress position and heading.
Figure 3: Global-Local Unity Feature (GLUF). A VGG-16 extracts feature map F from input image I , a Non-local Block produces semi-global feature map X composed of local descriptors xp . Two clustering branches compute cluster weights and similarities with learnable centers ak to obtain feature Xk,p∗ , which are aggregated across spatial sites and concatenated to form the GLUF vector u .
Figure 4: Bearing-Naver’s operating mode. At each step, we first collect the UVP and its corresponding RST features, use Bearing-UAV to regress the position and heading, and then compute the azimuth from the current position to the next waypoint, aligning the heading with this azimuth to update UAV state for the next step.
Datasets
Uni-1652
SUES
Dense
GTA-UAV
Ours
#UAV
37,854
24,210
18,198
33,763
90k × 2
#Satellite
1,652
200
9,096
9,096
1,024
Scene
Building
Campus
Campus
City
Multi-city
GeoTag
Fine
Coarse
Fine
Synthetic
Fine
Contiguous *
✗
✗
✓
✓
✓
Unaligned
✗
✗
✗
✓
✓
Table 1: Comparison between Bearing-UAV-90K and other geo-localization datasets. Existing datasets rarely consider unaligned scenarios. Ours focuses on arbitrary rotations (with varying IoUs and challenging misalignment) between UVPs and RSTs, and provides heading-aware annotations, forming a more comprehensive testbed for UAV localization and navigation. SUES = SUES-200, Dense = DenseUAV, Contiguous * : RSTs form a contiguous map.
Geo-Localization
Navigation
Method
Recall@1 ↑
LSR@15 ↑
HSR@15 ↑
MLE ↓
MHE ↓
SR@20 ↑
SPL ↑
NE ↓
Sat.
UAV
Sat.
UAV
Sat.
UAV
Sat.
UAV
Sat.
UAV
Sat.
UAV
Sat.
UAV
Sat.
UAV
University-1652 [ 52 ]
63.29
60.20
16.31
15.11
–
–
31.78
33.15
–
–
0.00
0.00
0.00
0.00
546.33
602.96
SUES-200 [ 55 ]
71.96
66.60
16.78
15.76
–
–
29.09
30.83
–
–
0.00
0.00
0.00
0.00
439.50
618.85
DenseUAV [ 6 ]
80.04
73.43
17.23
16.54
–
–
26.82
28.79
–
–
0.00
0.00
0.00
0.00
627.64
651.93
GTA-UAV [ 18 ]
76.99
70.71
31.33
27.96
–
–
25.24
28.43
–
–
0.00
0.00
0.00
0.00
666.51
661.91
Table 2: The geo-localization and navigation performance in satellite views and UAV views (U-S cross-view). WA: weather augmentation.
Figure 5: Localization and heading performance both consistently improve with increasing dataset scale and gradually begin to saturate once the dataset scale exceeds 60% of Bearing-UAV-90K.
Cities
Recall@1 ↑
LSR@15 ↑
MLE ↓
MHE ↓
Sat.
UAV
Sat.
UAV
Sat.
UAV
Sat.
UAV
A
89.43
79.92
97.20
80.36
6.14
10.29
4.92
12.07
B
88.76
83.39
96.36
88.14
6.44
8.97
4.28
10.27
C
90.40
85.43
98.22
88.01
5.54
8.67
4.00
17.15
D
88.18
78.90
96.18
82.72
6.62
10.20
4.90
15.71
AD
89.58
80.54
97.96
83.14
6.03
10.00
4.97
14.65
Table 3: Effect of multi-city diversity on model generalization performance. We design four distinct city combinations for training.
Weather
LSR@15 ↑
HSR@15 ↑
MLE ↓
MHE ↓
Sat.
UAV
Sat.
UAV
Sat.
UAV
Sat.
UAV
Rain
95.63
89.75
93.65
79.06
6.70
8.41
6.15
12.04
Snow
94.80
88.79
94.33
79.28
6.95
8.64
6.04
12.04
Fog
95.61
89.02
93.86
78.72
6.68
8.55
6.36
12.56
Bright
98.02
92.25
97.59
83.71
5.72
7.69
4.68
10.31
Mixed
96.34
90.33
95.44
81.29
6.35
8.19
5.59
11.24
Table 4: Weather robustness. Rows 1–6: augmented model under six conditions; Baseline: no augmentation under normal weather.
Figure 6: Navigation performance comparison. The purple dashed line here denotes the predefined navigation trajectory #1 in City D. ■ start, ★ end, × failure, ▲ success. The corresponding UAV-view frames across the 45 navigation steps are shown in Suppl. Fig. 16 .
Metric
Uni-1652
SUES
Dense
GTA-UAV
Ours
MS ↓ (MB)
102
658
86
326
66
IT ↓ (ms)
10.1
1907.5
593.6
246.5
133.5
GFLOPs ↓
16.4 @ 256
98.7 @ 384
8.5 @ 256
20.1 @ 384
10.5 @ 256
Table 5: Model efficiency. MS: model size; IT: inference time. GFLOPs are computed at 256 × 256 or 384 × 384 as in prior work.
GLUF
RCE
PSG
CA
R@1 ↑
LSR ↑
HSR ↑
MLE ↓
MHE ↓
✗
✓
✓
✓
67.19
57.29
61.10
14.90
21.62
✓
✗
✓
✓
82.66
87.77
70.80
8.93
14.99
✓
✓
✗
✓
83.97
88.68
75.28
8.67
13.44
✓
✓
✓
✗
83.70
89.06
76.68
8.59
13.20
✓
✓
✓
✓
83.17
89.36
77.21
8.61
12.90
Table 6: Ablation study of the Bearing-UAV on the UAV-view data.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Type
Acronym
Description
Concept
UAV
Unmanned Aerial Vehicle
GNSS
Global Navigation Satellite System
CVGL
Cross-View Geo-Localization
M2T
Match-to-Tile
Data
UVP
UAV-View Patch
RST
Remote-Sensing Tile
Appendix
Table 7: List of main acronyms used in this paper.
Method
Recall@1 ↑
LSR@15 ↑
HSR@15 ↑
MLE ↓
MedLE ↓
MHE ↓
MedHE ↓
Sat.
UAV
Sat.
UAV
Sat.
UAV
Sat.
UAV
Sat.
UAV
Sat.
UAV
Sat.
UAV
Ours ResNet18
83.88
75.34
88.38
71.83
73.20
46.83
8.83
12.09
7.89
10.56
14.59
26.52
8.54
16.16
Ours ViT-Small
86.38
81.39
93.35
85.47
77.58
51.57
7.53
9.46
6.73
8.02
12.02
24.89
7.69
14.45
Ours Mobile-V3S
86.41
79.76
93.52
81.20
83.25
64.94
7.49
10.34
6.73
8.72
10.57
19.53
6.28
10.14
Ours VGG-16
90.76
83.17
98.33
89.36
98.12
77.21
5.66
8.61
5.05
7.30
4.15
12.90
2.96
7.20
Appendix
Table 8: Comparison of localization and heading performance with four different backbones in both satellite views and UAV views (U-S cross-view). Bearing-UAV with the VGG-16 backbone achieves the best overall results. Mobile-V3S = MobileNet-V3-Small.
Figure 7: Satellite images of four cities with distinct landscapes. Below the satellite images are their selected U-S cross-view image pairs (UAV views left). City A is dominated on the right side by vegetation; City B consists mainly of densely packed low-rise buildings; City C contains many tall buildings and a wide river; City D lies in mountainous terrain with sparse, highly similar buildings. They exhibit prominent cross-view appearance gaps, illumination variations, and scene changes, covering diverse landforms such as buildings, roads, forests, mountains, and rivers.
Figure 8: Two examples of UAV-satellite cross-view UVP sampling during dataset construction. In each row, the left panel shows the sampled patch in the RSB, and the middle and right panels show the corresponding satellite-view patch and UAV-view patch.
Figure 9: Localization and orientation performance under different city combinations.
Figure 10: Four types of weather augmentations. The first row shows four original cross-view sample pairs from the four cities, where the green and blue boxes denote the UVP and its corresponding satellite-view patch. The four rows below visualize the corresponding illumination, fog, rain, and snow augmentations.
Figure 11: Effect of Weather Augmentation on Bearing-UAV. Mixed-weather training further boosts and stabilizes performance compared with the non-augmented setting.
Figure 12: Localization and heading error distributions and scatter plots of 9k unseen test samples in satellite and UAV views.
Figure 13: Heatmaps of mean localization and heading errors aggregated per RSB in two views across four cities.
Figure 14: Visualization of U-S cross-view localization and heading errors on real satellite images from four cities. The four satellite images in the top-left, top-right, bottom-left, and bottom-right panels correspond to City A–D , respectively, and overlay the localization and heading results of all unseen test samples under cross-view setting. In each image, blue/red dots denote the ground-truth/predicted positions, while blue/red arrows denote the ground-truth/predicted heading vectors. The green line segment connecting the endpoints of the two arrows represents the position error, whose magnitude is annotated by green numerical values , and the purple numeric values denote the heading error; smaller values indicate better accuracy, and perfect overlap between the red and blue arrows corresponds to the ideal case. This visualization enables precise, fine-grained inspection of localization/heading errors under different scene types across all cities. The satellite images are divided into fine-grained RSTs, such that the surrounding terrain of any sample can be examined and analyzed in detail against the real-world surroundings. For example, in the high-rise building areas ( e.g. , the lower-right of City A and the lower region of City C), in feature-sparse regions ( e.g. , forests in the upper-right of City A, green fields and the central river in City C, and open grounds in City D), as well as in dense clusters of similar buildings ( e.g. , City B and City D), many samples exhibit noticeably large localization and/or heading errors. These errors are likely due to the adverse effects of cross-view appearance discrepancies, feature sparsity, and feature similarity.
Figure 15: Navigation performance comparison on seven additional routes. We further provide a comprehensive comparison between Bearing-UAV (red trajectories) and four baseline methods on seven additional navigation routes across the four cities. The UAV step size is 25 m, and the waypoint arrival threshold is 20 m. All visual elements follow those in Fig. 6 of the main paper. The eight routes across Cities A–D range in length from 524 m to 1119 m. Each route contains 10–13 waypoints and covers diverse scene types, including buildings of varying scales, green areas, rivers, and playgrounds. Overall, Bearing-UAV exhibits the strongest adaptability to different urban layouts and highly winding paths. It successfully completes four of the eight routes and, for the remaining ones, still follows most of the designated path before failure. In contrast, baseline methods struggle in these complex scenes and along curved routes. The green GTA-UAV trajectories are often longer but show significant deviation from the reference path. The orange Uni-1652, yellow DenseUAV, and blue SUES-200 trajectories tend to drift away near the starting point, with some falling into local loops and circling within a small region. Notably, large positional deviations and flight drift mainly occur in vegetation areas (traj. #1 City A, traj. #2 City A, traj. #2 City D), similar building areas (traj. #1 City B, traj. #2 City D), and tall building areas (traj. #2 City C). These regions often suffer from feature sparsity, high feature similarity, and significant cross-view appearance discrepancies, all of which degrade localization and heading estimation performance.
Figure 16: Visualization of cross-view navigation frames of Bearing-UAV. We present a successful case for trajectory #1 in City D from Fig. 6 of the main paper. Frames are arranged top-to-bottom and left-to-right. In each step, the left panel shows the current RSB, where the blue rectangle indicates the UAV field of view and the airplane icon indicates the heading. The top-right inset shows the UVP, and the bottom-right inset reports outputs including heading angle, relative coordinates, direction vector, geographic coordinates, and the RSB index.
Figure 17: The UAV cross-view navigation process corresponding to trajectory #2 in City A, as shown in Fig. 15 (first row, middle). This is a more challenging case in which navigation eventually fails. However, the UAV still follows a substantial portion of the planned path before drifting away. We visualize overlapping portion (first 60 steps) to highlight cross-view discrepancies. Significant spatial differences between UVP and satellite patches arise from 3D parallax and illumination variations at low altitudes. Deviations occur at frames 26–34, likely due to waypoint turns and repetitive building patterns. Despite these challenges, Bearing-UAV remains robust along winding trajectories.
GLUF
RCE
PSG
CA
R@1 ↑
LSR ↑
HSR ↑
MLE ↓
MHE ↓
✗
✓
✓
✓
67.33
58.56
83.67
14.55
10.41
✓
✗
✓
✓
90.55
98.10
95.11
5.74
5.51
✓
✓
✗
✓
90.65
98.40
96.96
5.70
4.72
✓
✓
✓
✗
89.88
98.28
96.30
5.68
5.12
✓
✓
✓
✓
90.76
98.33
98.12
5.65
4.15
Appendix
Table 9: Ablation study under satellite-view settings.
Cross-view geo-localization for Unmanned Aerial Vehicles (UAVs) operating in GNSS-denied environments remains challenging due to the severe geometric discrepancy between oblique UAV imagery and orthogonal satellite maps. Most existing methods address this problem through a decoupled pipeline of place retrieval and pose estimation, implicitly treating perspective distortion as appearance noise rather than an explicit geometric transformation. In this work, we propose a geometry-aware UAV geo-localization framework that explicitly models the 3D scene geometry to unify coarse place recognition and fine-grained pose estimation within a single inference pipeline. Our approach reconstructs a local 3D scene from multi-view UAV image sequences using a Visual Geometry Grounded Transformer (VGGT), and renders a virtual Bird's-Eye View (BEV) representation that orthorectifies the UAV perspective to align with satellite imagery. This BEV serves as a geometric intermediary that enables robust cross-view retrieval and provides spatial priors for accurate 3 Degrees of Freedom (3-DoF) pose regression. To efficiently handle multiple location hypotheses, we introduce a Satellite-wise Attention Block that isolates the interaction between each satellite candidate and the reconstructed UAV scene, preventing inter-candidate interference while maintaining linear computational complexity. In addition, we release a recalibrated version of the University-1652 dataset with precise coordinate annotations and spatial overlap analysis, enabling rigorous evaluation of end-to-end localization accuracy. Extensive experiments on the refined University-1652 benchmark and SUES-200 demonstrate that our method significantly outperforms state-of-the-art baselines, achieving robust meter-level localization accuracy and improved generalization in complex urban environments.
Haoyuan Li, Wen Yang, Fang Xu +4
School of Electronic Information, Wuhan University, Wuhan 430072, China · School of Artificial Intelligence, Wuhan University, Wuhan 430072, China · Technology and Engineering Center for Space Utilization and the Key Laboratory of Space Utilization, Chinese Academy of Sciences, Beijing 100094, China +1
Cross-view geo-localization (CVGL) is fundamental for precise localization and navigation in GPS-denied environments, aiming to match ground or UAV imagery with satellite views. Existing approaches often rely on global feature alignment, but they suffer from substantial domain shifts induced by varying regional textures and weather conditions. This issue becomes even more pronounced in UAV-based scenarios, where the broader perspective inevitably introduces dense, fine-grained objects, creating significant visual clutter. To address this, we draw inspiration from Object-Centric Learning (OCL) and propose InfoGeo, an information-theoretic framework designed to enhance robustness and generalization. InfoGeo reformulates the optimization as an information bottleneck process with two core objectives: (i) maximizing view-invariant information by aligning the object-centric structural relations across views, and (ii) minimizing view-specific noisy signals through cross-view knowledge constraints. Extensive evaluations across diverse benchmarks and challenging scenarios demonstrate that InfoGeo significantly outperforms state-of-the-art methods.
Hongyang Zhang, Maonan Wang, Ziyao Wang +2
School of Science and Engineering, The Chinese University of Hong Kong (Shenzhen), Guangdong 518172, China · Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Hong Kong SAR 999077, China · Shanghai AI Laboratory, Shanghai 200032, China
Cross-view UAV geolocalization is fundamentally a challenging large-scale image retrieval task, aiming to determine the geographic coordinates of Unmanned Aerial Vehicle (UAV) queries by matching them against an extensive geo-tagged satellite image database. Most existing methods learn separate feature representations for each view and determine the final prediction using naive heuristics to assess feature similarity, thereby neglecting to model the crucial cross-view relationships. In this paper, we propose SkyLink, a novel plug-and-play ranking framework that pioneers joint relational modeling of inter-view relationships to enhance cross-view UAV geolocalization. SkyLink leverages a Large Vision-Language Model (LVLM) to model the intricate visual-semantic relationships between UAV and satellite views, facilitating effective cross-view matching. To further refine the learning process, we introduce a relational-aware loss. It leverages soft labels to provide a more nuanced supervision signal, mitigating the harsh penalty on near-positive pairs. This approach enhances both training stability and the model's discriminative capacity. Extensive experiments conducted across multiple base retrieval architectures and benchmark datasets demonstrate that SkyLink significantly boosts the ranking effectiveness of existing models, consistently achieving superior performance in various challenging scenarios.
Bowen Liu, Pengyue Jia, Wanyu Wang +9
Department of Data Science, City University of Hong Kong, Hong Kong · Information Systems, City University of Hong Kong, Hong Kong · College of Computer Science and Technology, Zhejiang University of Technology, Zhejiang