Cross-view geo-localization（CVGL） is beset with numerous difficulties and challenges, mainly due to the significant discrepancies in viewpoint, the intricacy of localization scenarios, and global localization needs. Given these challenges, we present a novel cross-view image geo-localization framework. The experimental results demonstrate that the proposed framework outperforms existing methods on multiple public datasets and self-built datasets. To improve the cross-view geo-localization performance of the framework on a global scale, we have built a novel global cross-view geo-localization dataset, CV-Cities. This dataset encompassing a diverse range of intricate scenarios. It serves as a challenging benchmark for cross-view geo-localization.

We collected 223,736 ground images and 223,736 satellite images with high-precision GPS coordinates of 16 typical cities in five continents. To download this dataset, you can click: 🤗CV-Cities or 🤗CV-Cities (mirror).

				Colorbar
Capetown, South Africa	London, UK	Melbourne, Australia	Mexico city, Mexico
New York, USA	Paris, France	Rio, Brazil	Taipei, China
Losangeles, USA	Manila, Philipine	Santiago, Chile	Sydney, Australia
Seattle, USA	Singapore	Barcelona, Span	Tokyo, Japan

City scene		Nature scene
Water scene		Occlusion

London, UK	Rio, Brazil	Seattle, USA
Singapore	Sydney, Australia	Taipei, China

🚧 Under Construction

python train/train_cvcities.py

This code is based on the amazing work of:

• DINOv2
• Sample4Geo
• QGIS

  @ARTICLE{huangCVCities2024,
  author={Huang, Gaoshuang and Zhou, Yang and Zhao, Luying and Gan, Wenjian},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing}, 
  title={CV-Cities: Advancing Cross-View Geo-Localization in Global Cities}, 
  year={2025},
  volume={18},
  number={},
  pages={1592-1606},
  keywords={Urban areas;Satellite images;Accuracy;Feature extraction;Training;Location awareness;Buildings;Manuals;Benchmark testing;Transformers;Cross-view geo-localization (CVGL);dataset;global cities;image retrieval;visual place recognition},
  doi={10.1109/JSTARS.2024.3502160}}