OpenCVL: An Open, Diverse, and Large-Scale Dataset for Fine-Grained Cross-View Localization
Authors: Zimin Xia, Mubariz Zaffar, Junsheng Fu, Alexandre Alahi, Julian F. P. Kooij
Organizations: École Polytechnique Fédérale de Lausanne (EPFL), Switzerland · Southern University of Science and Technology (SUSTech), China · Delft University of Technology, The Netherlands · Zenseact
Fine-grained Cross-View Localization (CVL) estimates the precise position and orientation of a ground-level image by aligning it with geo-referenced aerial imagery, offering a scalable alternative to Global Navigation Satellite Systems (GNSS) in challenging urban environments. Existing datasets rely on data collected with high-end sensor suites, which inherently limit image diversity and scalability. While in-the-wild images are abundant, their noisy geo-tags make them unsuitable for reliable evaluation. To bridge this gap, we introduce OpenCVL, a large-scale, diverse, and open dataset containing 617,388 ground-aerial image pairs spanning 41 cities across four European countries. All images are sourced from permissive platforms, ensuring long-term accessibility and supporting open and reproducible research. The training set combines images captured with high-end sensors with diverse in-the-wild imagery. We further develop a data curation framework that filters and corrects pose annotations to construct reliable in-the-wild evaluation data. In addition, OpenCVL includes dedicated cross-area and snowy test sets to assess generalization and robustness. Experiments with a state-of-the-art CVL model on OpenCVL show that incorporating noisy in-the-wild data consistently improves performance on clean test sets, suggesting a promising direction for scaling CVL with diverse real-world imagery.
Figures & tables
Figure 1: Overview of OpenCVL , an Open dataset for fine-grained Cross-View Localization , spanning 41 cities across four European countries and covering more than 7,000 km 2 . The dataset encompasses diverse geographic regions and provides substantial variability in camera types, viewpoints, as well as temporal and weather conditions. The ground-level images are collected from Mapillary [ 19 ] (CC BY-SA) and ZOD [ 3 ] (CC BY-SA). The aerial imagery are sourced from open data from Lantmäteriet [ 11 ] (Sweden, CC BY 4.0), PDOK [ 20 ] (Netherlands, CC BY 4.0), Kartverket/Geovekst [ 10 ] (Norway, Åpne data, Norge digitalt-lisens), and GUGiK Geoportal [ 8 ] (Poland, INSPIRE noConditionsApply / noLimitations).
Sweden
Poland
Norway
Netherlands
Ground-level: ZOD (# images)
241107
29663
1403
2314
Ground-level: Mapillary (# images)
86540
117510
48103
90748
Aerial Image Resolution (m/pixel)
0.16
0.1
0.04–0.1
0.045
Table 1: Statistics of the proposed OpenCVL dataset across target countries.
Figure 2: Examples of pose verification in OpenCVL test sets curated from ZOD. We verify the projected camera poses and LiDAR point cloud (shown in purple) on the aerial imagery. Left: good alignment. Middle: misalignment, where the red boxes highlight the incorrect camera location and the mismatch between building edges and LiDAR points. Right: examples of selected snowy samples with good alignment.
Figure 3: Our framework for estimating accurate poses for Mapillary images. Feature matching is performed between an anchor ZOD image and the Mapillary image, followed by pose estimation using 2D–3D correspondences obtained from the globally registered ZOD LiDAR point cloud. The quality of the estimated pose is illustrated on the right by projecting the ZOD point cloud into the Mapillary image using our refined pose compared to the raw Mapillary GNSS pose.
Figure 4: Examples of diverse ground images and their corresponding aerial views in the OpenCVL dataset. Our dataset uniquely contains ground-level images taken by pedestrians and cyclists in truly in-the-wild conditions.
Dataset
Camera Type
Mounting
Temporal Diversity
Coverage
Data Source
VIGOR [ 43 ]
Fixed
Vehicle & backpack
Limited
4 U.S. cities
Restricted
FordAV [ 24 ]
Fixed
Vehicle
Diverse
Michigan, U.S.
Restricted
KITTI [ 24 ]
Fixed
Vehicle
Limited
Karlsruhe, DE
Restricted
RobotCar [ 37 ]
Fixed
Vehicle
Diverse
Oxford, UK
Restricted
OpenCVL
Mixed
Mixed
Diverse
4 EU countries
Open-access
Table 2: Comparison of existing fine-grained CVL datasets to our OpenCVL dataset.
Cross-area test set
Snowy test set
In-the-wild test set
Training data
Loc. (m)
Ori. ( ∘ )
Loc. (m)
Ori. ( ∘ )
Loc. (m)
Ori. ( ∘ )
Mean
Med.
Mean
Med.
Mean
Med.
Mean
Med.
Mean
Med.
Mean
Med.
KITTI w. ori. prior
20.07
18.14
80.65
80.49
17.30
15.62
79.75
65.04
17.12
15.15
86.25
85.03
KITTI w/o. ori. prior
15.57
13.60
79.75
72.93
15.92
14.48
83.30
76.41
15.98
14.07
70.72
57.15
OpenCVL (ZOD only)
7.55
4.50
16.20
5.74
8.04
5.74
8.87
3.26
8.61
6.85
35.59
12.45
OpenCVL (w/o. weig.)
6.87
4.78
13.01
6.34
8.04
6.05
10.91
5.34
8.16
6.47
25.87
12.93
Table 3: Evaluation of Loc 2 [ 39 ] on the OpenCVL test sets under different training settings. We compare models trained on KITTI with and without the orientation prior, training on OpenCVL using only ZOD data, joint training with additional Mapillary data without loss downweighting, and the final setting that applies a global downweighting factor of 0.3 to Mapillary samples in OpenCVL. Best in bold.
Figure 5: Orientation error histograms on the KITTI cross-area test split.
Training data
Mean Loc. (m)
Med. Loc. (m)
Mean Ori. ( ∘ )
Med. Ori. ( ∘ )
KITTI (w/o ori. prior)
11.71
9.11
55.18
33.41
OpenCVL (ZOD only)
12.04
10.34
75.00
59.54
OpenCVL
11.25
9.40
71.17
46.96
Table 4: KITTI cross-area evaluation without an orientation prior. For localization, the model trained on our full OpenCVL dataset generalizes slightly better to KITTI cross-area test set than a model trained on the KITTI distribution itself. The results “KITTI (w/o ori. prior)” are taken from the original Loc 2 paper [ 39 ] . Best in bold.
Figure 6: Qualitative results of Loc 2 across different test sets (green/yellow arrow: true/predicted pose). Lines show Loc 2 ’s top 20 correspondences by matching score.
Finetuning labels
Mean Loc. (m)
Med. Loc. (m)
Mean Ori. ( ∘ )
Med. Ori. ( ∘ )
No finetuning
8.05
5.75
26.73
10.57
Raw GNSS pose
7.41
5.73
22.09
7.70
OpenSfM pose
7.20
5.23
23.44
9.27
Our corrected pose
6.92
5.14
17.98
7.01
Table 5: The impact of different pose labels for Mapillary images. Best in bold.
Figure 7: Visual comparisons of the different pose annotations available for Mapillary images in our OpenCVL evaluation. The corrected Mapillary pose is significantly better than the raw Mapillary GNSS tag and the Mapillary-reported OpenSfM corrected pose.
Same-area
Cross-area
Training strategy
↓ Localization (m)
↓ Orientation ( ∘ )
↓ Localization (m)
↓ Orientation ( ∘ )
Mean
Med.
Mean
Med.
Mean
Med.
Mean
Med.
KITTI only
1.13
0.77
1.97
1.43
5.60
3.01
3.32
2.12
OpenCVL + KITTI
1.12
0.77
1.90
1.37
5.43
2.99
4.06
2.38
OpenCVL → KITTI
0.93
0.61
1.62
1.19
5.07
2.84
3.30
2.15
Table 6: KITTI evaluation with a ±10∘ orientation prior. Best in bold.
Country
City
Samples
NL
Amsterdam
21664
NL
Breda
4848
NL
Delft
2174
NL
Eindhoven
6774
NL
Groningen
5626
NL
Haarlem
2750
Table 7: Number of samples per city in the dataset.
Figure 8: Example of the homography-based preprocessing used for HC-Net.
Cross-area test set
Snowy test set
In-the-wild test set
Loc.
Ori.
Loc.
Ori.
Loc.
Ori.
Mean
Med.
Mean
Med.
Mean
Med.
Mean
Med.
Mean
Med.
Mean
Med.
HC-Net
13.87
13.96
99.73
106.08
12.82
12.85
94.89
107.73
12.66
12.28
85.89
88.27
CCVPE
11.95
7.64
25.59
4.56
10.81
7.66
7.41
2.48
16.24
12.51
38.64
7.88
Loc 2
6.72
4.30
11.52
5.24
7.80
5.48
5.09
3.52
7.90
6.19
26.07
12.01
Table 8: Extra baseline results. Best in bold.
Figure 9: Visualization of ZOD LiDAR pointcloud projected into the Mapillary images using the raw Mapillary pose, the OpenSfM pose, and our corrected pose. The structure in LiDAR pointcloud overlaps well with the structure in the images when projected using our corrected pose for the Mapillary images.