Home

Awesome

VIGOR : Cross-View Image Geo-localization beyond One-to-one Retrieval

<img width=913 height=300 src="data/Architecture.jpg"/>

This repository provides the dataset and code used in "VIGOR : Cross-View Image Geo-localization beyond One-to-one Retrieval".

@inproceedings{zhu2021vigor,
  title={VIGOR: Cross-View Image Geo-localization beyond One-to-one Retrieval},
  author={Zhu, Sijie and Yang, Taojiannan and Chen, Chen},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={3640--3649},
  year={2021}
}

Dataset

Please follow the guideline to download and prepare the dataset. For the resolution issue, authors of SliceMatch has measured the resolution of all cities and revised the label. If you want to use this label, you may find it at their github repo. <img width=781 height=300 src="data/collection.jpg"/>

Requirement

- Python >= 3.5, Opencv, Numpy, Matplotlib
- Tensorflow == 1.13.1

Evaluation from npy

Download the same-area models and npy files from the link, unzip (tar -zxvf) it in "./data/". Then run the script:

python evaluate_from_npy.py

Training and evaluating from model

Download the initialization weights from ImageNet, put it in "./data/". Then run the script to train a simple SAFA baseline:

python train_SAFA.py

Run the script to train with our method:

python train_overall.py

Reference

- https://github.com/shiyujiao/cross_view_localization_SAFA
- https://github.com/Jeff-Zilence/Explain_Metric_Learning
- https://github.com/david-husx/crossview_localisation.git