MEAN 2024 [paper][model] [Cite]

Multilevel Embedding and Alignment Network With Consistency and Invariance Learning for Cross-View Geo-Localization

Zhongwei Chen, Zhaoxu Yang*, Haijun Rong*

School of Aerospace Engineering, Xi'an Jiaotong University China
State Key Laboratory for Strength and Vibration of Mechanical Structures
Shaanxi Key Laboratory of Environment and Control for Flight Vehicle

This repository is the official implementation of the paper "Multilevel Embedding and Alignment Network With Consistency and Invariance Learning for Cross-View Geo-Localization" (https://arxiv.org/abs/2412.14819).

The current version of the repository can cover the experiments reported in the paper, for researchers in time efficiency. And we will also update this repository for better understanding and clarity.

🔥 News

• [May 17, 2025]: MEAN is accepted by TGRS'25 🎉
• [Apr 15, 2025]: We uploaded the visualization code.
• [Feb 28, 2025]: We released the MEAN model and its pre-trained weights.

---

📚 Table of contents

• Dataset Access
• Dataset Structure
• Train and Test
• Pre-trained Checkpoints
• License
• Acknowledgments
• Citation

💾 Dataset Access

Please prepare University-1652, SUES-200

📁 Dataset Structure

University-1652 Dataset Directory Structure

├── University-1652/
│   ├── train/
│       ├── drone/                   /* drone-view training images 
│           ├── 0001
|           ├── 0002
|           ...
│       ├── satellite/               /* satellite-view training images       
│   ├── test/
│       ├── query_drone/  
│       ├── gallery_drone/  
│       ├── query_satellite/  
│       ├── gallery_satellite/ 

SUES-200 Dataset Directory Structure

├─ SUES-200
  ├── Training
    ├── 150/
    ├── 200/
    ├── 250/
    └── 300/
  ├── Testing
    ├── 150/
    ├── 200/ 
    ├── 250/	
    └── 300/

🚀 Train and Test

For University-1652 Dataset

Train: run train_university.py, with --only_test = False.

Test: run train_university.py, with --only_test = True, and choose the model in --ckpt_path.

For SUES-200 Dataset

Train: run train_SUES-200.py, with --only_test = False.

Test: run train_SUES-200.py, with --only_test = True, and choose the model in --ckpt_path.

🤗 Pre-trained Checkpoints

We provide the trained models in the link below:

Baidu Netdisk Link: [https://pan.baidu.com/s/1QoYcr2XXy5z0oFh2Tzi40A?pwd=6666 提取码: 6666]

Google Drive Link: [https://drive.google.com/drive/folders/13aFkUDNzqOHAvDfaloh14RMvOuPZqi3G?usp=drive_link]

We will update this repository for better clarity ASAP, current version is for quick research for researchers interested in the cross-view geo-localization task.

🎫 License

This project is licensed under the Apache 2.0 license.

📌 Citation

If you find this code useful for your research, please cite our papers.

@article{chen2024multi,
  author={Chen, Zhongwei and Yang, Zhao-Xu and Rong, Hai-Jun},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Multi-Level Embedding and Alignment Network with Consistency and Invariance Learning for Cross-View Geo-Localization}, 
  year={2025},
  volume={63},
  number={},
  pages={1-15}

🙏 Acknowledgments

This repository is built using the Sample4Geo, MCCG and DAC repositories. Thanks for their wonderful work.