Abstract

Cross-view geolocalization identifies the geographic location of street view images by matching them with a georeferenced satellite database. Significant challenges arise due to the drastic appearance and geometry differences between views. In this paper, we propose a new approach for cross-view image geo-localization, i.e., the Panorama-BEV Co-Retrieval Network. Specifically, by utilizing the ground plane assumption and geometric relations, we convert street view panorama images into the BEV view, reducing the gap between street panoramas and satellite imagery. In the existing retrieval of street view panorama images and satellite images, we introduce BEV and satellite image retrieval branches for collaborative retrieval. By retaining the original street view retrieval branch, we overcome the limited perception range issue of BEV representation. Our network enables comprehensive perception of both the global layout and local details around the street view capture locations. Additionally, we introduce CVGlobal, a global cross-view dataset that is closer to real-world scenarios. This dataset adopts a more realistic setup, with street view directions not aligned with satellite images. CVGlobal also includes cross-regional, cross-temporal, and street view to map retrieval tests, enabling a comprehensive evaluation of algorithm performance.

News

• The code for Street View-BEV Co-retrieval inference is now available. If there is any missing code or abnormality, you can report it to me in the issue.
• The training and testing code for the BEV branch on CVACT has been released.
• Source code of BEV transformation is released（CVACT/CVUSA）.

Installation

Clone this repo to a local folder:

git clone https://github.com/yejy53/EP-BEV.git
cd EP-BEV

Environment Setup

conda create -n EP-BEV python=3.9 -y
conda activate EP-BEV
pip install -r requirements.txt

If huggingface cannot download the weights successfully, you can add export HF_ENDPOINT="https://hf-mirror.com" at the end of .bashrc and reactivate it.

Data Preparation

The publicly available datasets used in this paper can be obtained from the following sources:

Preparing CVUSA Dataset. The dataset can be downloaded here.

Preparing CVACT Dataset. The dataset can be downloaded here.

Preparing VIGOR Dataset. The dataset can be downloaded here.

Preparing CVGlobal Dataset. The dataset can be downloaded here.

Data Structure:

├─ CVACT
  ├── ACT_data.mat
  ├── ANU_data_small/
    ├── bev/
    ├── satview_polish/ 
    ├── streetview/	
  └──ANU_data_test/

Use our pre-trained model for retrieval

1. You can download a pre-trained model (e.g. cvact) from huggingface and place it in ckpt folder.
2. You need to organize the generated BEV images into the above dataset format. You can download the generated BEV images directly from the following huggingface link to get consistent results, or generate BEV images yourself and then retrain.
3. When performing Street View-BEV Co-Retrieval, you only need to add the similarity of using a pure Street View image to the similarity of using a BEV image. The weights for using Street View search can be obtained from the following huggingface link. The method and weights for using Street View search also can be found in the Sample4G.

❤️ Acknowledgements

Our code is built on top of Sample4G and Boosting3DoF. We appreciate the previous open-source works.

BibTeX 🙏

If you have any questions, be free to contact with me!

@inproceedings{ye2025cross,
  title={Cross-view image geo-localization with Panorama-BEV Co-Retrieval Network},
  author={Ye, Junyan and Lv, Zhutao and Li, Weijia and Yu, Jinhua and Yang, Haote and Zhong, Huaping and He, Conghui},
  booktitle={European Conference on Computer Vision},
  pages={74--90},
  year={2025},
  organization={Springer}
}