UAV-GeoLoc

This repo contains the large-vocabulary datasets for our paper UAV-GeoLoc: A Large-vocabulary Dataset and Geometry-Transformed Method for UAV Geo-Localization. Given a top-down UAV image, retrieve a corresponding satellite image patch to infer the UAV's location.

📦 Usage && Dataset Structure

⬇️ Download Instructions

You need to download all parts of a category (e.g., all Country.zip.00X files) before extraction.

1. Clone the repository with Git LFS enabled:

git lfs install
git clone https://huggingface.co/datasets/RingoWRW97/UAV-GeoLoc

2.combine and extract the files

# For Country
cat Country.zip.* > Country.zip
unzip Country.zip

# For Terrain
cat Terrain.zip.* > Terrain.zip
unzip Terrain.zip

# For Rot
unzip Rot.zip

🖼️ Dataset Structure

Each folder under Country or Terrain (e.g., USA, Italy, Japan, etc.) contains N scenes for that region. Each scene is structured as follows:
Country/
├── Australia/
├────City(Sydney)
├──────Region
├──────── DB / (Satellite Map)
├──────── query / (24 dir,including height 100:150:25, heading 0:360:45)
├──────── semi_positive.json
├──────── positive.json
├── Brazil/
├── USA/
└── ...

🗂️ Index.zip (Train/Val/Test Splits)

The dataset includes a compressed file Index.zip that contains various .txt files used to define training, validation, and test splits across different components of the dataset.
After extracting Index.zip, the structure looks like: Index/
├── train.txt
├── train_all.txt
├── train_country.txt
├── train_db.txt
├── train_db_all.txt
├── train_db_country.txt
├── train_query.txt
├── train_query_all.txt
├── train_query_country.txt
├── train_query_test.txt
├── val.txt
├── val_all.txt
├── val_country.txt
├── val_db.txt
├── val_db_country.txt
├── val_query.txt
├── val_query_country.txt
├── test.txt
├── test_all.txt
├── test_country.txt
├── test_db.txt
├── test_query.txt

Each file defines a specific subset of the dataset used for:

*_query.txt: UAV query images
*_db.txt: Reference DB images
*_country.txt: only train on country class
*.txt: only train on terrain class
*_all.txt: Union of all images in a given category

📸 Result on Rot

‘Fire’ denotes results trained on our proposed dataset. ‘Box’ indicates that the model is trained with the LPN method.

📄 License

This dataset is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).
Any modification of the dataset is strictly prohibited. The imagery was collected using Google Earth Studio, and appropriate attribution to Google must be provided in any derivative work or publication.