OpenCVL
EPFL, SUSTech, TU Delft, and Zenseact logos

ECCV 2026 (Spotlight) Sep 11, 2026 · Spotlight Presentation 9:25 AM CEST · Poster Session 3, #141, ExHall Poster, 10:30 AM CEST

OpenCVL: An Open, Diverse, and Large-Scale Dataset for Fine-Grained Cross-View Localization

Zimin Xia1 2 * †, Mubariz Zaffar3 *, Junsheng Fu4, Alexandre Alahi1, and Julian F. P. Kooij3

1 École Polytechnique Fédérale de Lausanne (EPFL), Switzerland 2 Southern University of Science and Technology (SUSTech), China 3 Delft University of Technology (TU Delft), The Netherlands 4 Zenseact *Equal contribution, †corresponding author

Download Samples Access Dataset Developer Tools Paper Presentation
4
countries
41
cities
7,000+ km2
coverage
617,388
ground-aerial pairs

Open, diverse, and large-scale data for fine-grained CVL.

Open

Ground imagery comes from ZOD and Mapillary, paired with aerial imagery from national mapping agencies with open access.

Diverse

The dataset combines calibrated vehicle-mounted sensors with in-the-wild images from pedestrians, cyclists, cars, and other capture platforms.

Large-scale

OpenCVL covers four countries, 41 cities, more than 7,000 km2, and 617,388 ground-aerial image pairs.

Country-level data statistics and aerial resolution.

OpenCVL spans Sweden, the Netherlands, Poland, and Norway. Each ground-level image is paired with a 100 m by 100 m aerial crop from national open mapping sources.

OpenCVL coverage map across Sweden, Norway, the Netherlands, and Poland with sample ground and aerial images.
Sweden 327,647 ground-aerial pairs
ZOD
241,107
Mapillary
86,540
Resolution
0.16 m/pixel
Netherlands 93,062 ground-aerial pairs
ZOD
2,314
Mapillary
90,748
Resolution
0.045 m/pixel
Poland 147,173 ground-aerial pairs
ZOD
29,663
Mapillary
117,510
Resolution
0.1 m/pixel
Norway 49,506 ground-aerial pairs
ZOD
1,403
Mapillary
48,103
Resolution
0.04-0.1 m/pixel

Reliable supervision plus real-world diversity.

ZOD ground imagery

Vehicle-mounted front-facing cameras, LiDAR, and high-end GNSS provide accurate pose supervision for training and curated evaluation.

274,487 images

Mapillary imagery

Crowd-sourced street-level imagery adds broad variation in camera type, mounting, viewpoint, time, weather, and scene content.

342,901 images

Open aerial sources

Aerial crops are retrieved from national open-data mapping sources for Sweden, the Netherlands, Poland, and Norway.

0.04-0.16 m/pixel
Diverse OpenCVL examples with ground images above their corresponding aerial views.
Diverse ground-level captures paired with aerial views: pedestrian, cyclist, snowy, urban, night, and road scenes.

Splits for scale, generalization, seasons, and in-the-wild robustness.

Training split

Large-scale training

Combines ZOD and Mapillary ground imagery to scale supervised learning with both accurate pose supervision and broad viewpoint diversity.

Primary goal
Large-scale training
Ground source
ZOD + Mapillary

Dataset release and project resources.

Download Samples Preview a small OpenCVL sample package.
Access Dataset Register, then continue to the full dataset archives.
Developer Tools OpenCVL code and development tools.
Paper ECCV 2026 spotlight paper.

Citation

@inproceedings{opencvl2026,
  title = {OpenCVL: An Open, Diverse, and Large-Scale Dataset for Fine-Grained Cross-View Localization},
  author = {Xia, Zimin and Zaffar, Mubariz and Fu, Junsheng and Alahi, Alexandre and Kooij, Julian F. P.},
  booktitle = {European Conference on Computer Vision},
  year = {2026}
}