DAC

This repository is the official implementation of the paper "Enhancing Cross-View Geo-Localization with Domain Alignment and Scene Consistency".

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.

1. 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.

2. For SUES-200 dataset.

You need to split the origin dataset into the appropriate format using the script "DAC-->sample4geo-->dataset-->SUES-200-->split_datasets.py".

The processed format should be:

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

The train and test operation is similar to the University-1652 dataset but with the script train_sues200.py

3. Multi-weather University-1652 dataset

This part has not been well presented, but you can manually use it in the data augmentation script "DAC-->sample4geo-->dataset-->university.py". We provide multi-weather augmentation settings using python package "albumentations", you can add them into the transforms to evaluate the model on Multi-weather Univerisy-1652 dataset.

4. Models

We provide the trained models in the link below: https://drive.google.com/file/d/140xgtckQkRwqgszD1wWiDH6lV7xufa8r/view?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.

5. Acknowledgement

This repository is built using the Sample4Geo[https://github.com/Skyy93/Sample4Geo] and MCCG[https://github.com/mode-str/crossview] repositories.