GeoBridge: A Semantic-Anchored Multi-View Foundation Model Bridging Images and Text for Geo-Localization

Zixuan Song^1,3, Jing Zhang^{2,3 †}, Di Wang^{2,3 †}, Zidie Zhou¹, Wenbin Liu¹, Haonan Guo^{2,3 †}, En Wang^{1 †}, Bo Du^{2,3 †}.

¹ Jilin University, ² Wuhan University, ³ Zhongguancun Academy.

^† Corresponding author

Update | Abstract | Datasets | Model | Usage | Statement

🔥 Update

2026.05.14

• The dataset is now available.

2026.04.08

• The model is now available.

2026.03.26

• Code is now available.

2026.02.21

• The paper is accepted by CVPR 2026! 🎉

2025.12.03

• The paper is post on arXiv! (arXiv GeoBridge)

🌞 Abstract

Cross-view geo-localization infers a location by retrieving geo-tagged reference images that visually correspond to a query image. However, the traditional satellite-centric paradigm limits robustness when high-resolution or up-to-date satellite imagery is unavailable. It further underexploits complementary cues across views (e.g., drone, satellite, and street) and modalities (e.g., language and image). To address these challenges, we propose GeoBridge, a foundation model that performs bidirectional matching across views and supports language-to-image retrieval. Going beyond traditional satellite-centric formulations, GeoBridge builds on a novel semantic-anchor mechanism that bridges multi-view features through textual descriptions for robust, flexible localization. In support of this task, we construct GeoLoc, the first large-scale, cross-modal, and multi-view aligned dataset comprising over 50,000 pairs of drone, street-view panorama, and satellite images as well as their textual descriptions, collected from 36 countries, ensuring both geographic and semantic alignment. We performed broad evaluations across multiple tasks. Experiments confirm that GeoLoc pre-training markedly improves geo-location accuracy for GeoBridge while promoting cross-domain generalization and cross-modal knowledge transfer.

Figure 1. Schematic diagram of GeoBridge.

Figure 2. Overall workflow.

📖 Datasets

GeoLoc Dataset

The GeoLoc dataset is a large-scale multi-view and cross-modal geo-localization dataset, consisting of aligned drone, street-view, and satellite imagery.

Data Sources

• Drone images are collected from OpenAerialMap, a public platform that provides access to openly licensed aerial and UAV imagery.

• The corresponding street-view panoramas and satellite images are obtained from Google Street View and Google Earth / Google Maps Platform services.

Data Release Statement

The drone imagery in our dataset is derived from publicly available and open-access sources.
We will release the processed drone-image subset used in this project after dataset organization is completed.

The processing pipeline, filtering strategy, and cross-view alignment details are described in the main paper and the supplementary material. We release the processed drone-image subset required for training and testing at Hugging Face GeoLoc.

Due to the usage policies and service terms associated with Google Street View and related Google Maps Platform services, we cannot directly redistribute the street-view and satellite images collected from Google services. Therefore, these images will not be included in the public dataset release. Researchers who wish to reconstruct the complete multi-view data may use the official Google APIs, subject to Google’s applicable policies and terms.

🚀 Model

The GeoBridge model is now available on Hugging Face: Son12s/GeoBridge. Please visit the repository page for download and usage details.

🔨 Usage

Data Preparation

Please organize the dataset as follows:

data/
├── train/
│   ├── drone/
│   ├── satellite/
│   └── street/
├── val/
│   ├── drone/
│   ├── satellite/
│   └── street/
└── test/
    ├── drone/
    ├── satellite/
    └── street/

Checkpoints

Please download the pretrained checkpoints and place them under:

checkpoints/
├── opts.yaml
└── best_net.pth

Evaluation

Supported evaluation settings include:

• drone ↔ satellite retrieval
• street ↔ satellite retrieval
• satellite ↔ street retrieval
• text → image retrieval

Example Tasks

GeoBridge supports the following tasks:

1. Cross-view geo-localization
   Retrieve geographically matched reference images across different views.

2. Bidirectional image retrieval
   Perform retrieval between drone, satellite, and street-view imagery.

3. Language-to-image retrieval
   Use natural language descriptions to retrieve semantically aligned geo-images.

🍭 Results

⭐ Citation

If you find GeoBridge helpful, please give a ⭐ and cite it as follows:

@misc{song2025geobridgesemanticanchoredmultiviewfoundation,
      title={GeoBridge: A Semantic-Anchored Multi-View Foundation Model Bridging Images and Text for Geo-Localization}, 
      author={Zixuan Song and Jing Zhang and Di Wang and Zidie Zhou and Wenbin Liu and Haonan Guo and En Wang and Bo Du},
      year={2025},
      eprint={2512.02697},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.02697}, 
}

🎺 Statement

For any other questions please contact Zixuan Song at jlu.edu.cn or gmail.com.