README.md

February 19, 2026 · View on GitHub

Drone Referring Localization: An Efficient Heterogeneous Spatial Feature Interaction Method For UAV Self-Localization

This repository contains code and dataset for the paper titled Drone Referring Localization: An Efficient Heterogeneous Spatial Feature Interaction Method For UAV Self-Localization.

News

  • 2024/8/28: Our dataset and code are released.

Table of contents

About Dataset

The dataset split is as follows:

SubsetUAV-viewSatellite-viewuniversities
Train6,7686,76810
test2,33127,9724

More detailed file structure:

├── UL14/
│   ├── train/
│       ├── PlaceName_Height_Index/                  
│           ├── UAV
│               ├── 0.JPG
│           ├── Satellite/              
│               ├── 0.tif
|         ...
│   ├── val/
        ├── PlaceName_Height_Index/
            ├── UAV
│               ├── 0.JPG
│           ├── Satellite/       
│               ├── 0.jpg
|               ├── 1.jpg
|               ├── 2.jpg
|               ...
|               ├── 11.jpg
│           GPS_info.json           /* UAV position in satellite images
|           label.json              /* Supplementary information such as latitude and longitude, mapsize
│   ├── test/                       /* Structure is same as val

Prerequisites

  • Python 3.7+
  • GPU Memory >= 8G
  • Numpy 1.26.0
  • Pytorch 2.0.0+cu118
  • Torchvision 0.15.0+cu118

Installation

It is best to use cuda version 11.8 and pytorch version 2.0.0. You can download the corresponding version from this website and install it through pip install. Then you can execute the following command to install all dependencies.

pip install -r requirments.txt

Create the directory for saving the training log and ckpts.

mkdir checkpoints

Dataset & Preparation

UL14 has been released in link.

Additionally, download the pretrain weight of cvt13 from this link.

Important: change the pretrain_path, train_dir, val_dir, test_dir in the config file.

Train & Evaluation

Training and Testing

Execute the following command to implement the entire process of training and testing.

bash train_test_local.sh

The setting of parameters in train_test_local.sh can refer to Get Started.

Evaluation

The following commands are required to evaluate MA@K and RDS.

cd checkpoints/<name>
python test_meter.py --config <name>

the <name> is the dir name in your training setting, you can find in the checkpoints/.

We also provide the baseline checkpoints, link.

unzip <file.zip> -d checkpoints
cd checkpoints/baseline
python test.py --test_dir <dataset_root>/test
python test_meter.py --config <name>

License

This project is licensed under the Apache 2.0 license.

Citation

The following paper uses and reports the result of the baseline model. You may cite it in your paper.

@misc{drl,
      title={Drone Referring Localization: An Efficient Heterogeneous Spatial Feature Interaction Method For UAV Self-Localization}, 
      author={Ming Dai and Enhui Zheng and Zhenhua Feng and Jiahao Chen and Wankou Yang},
      year={2024},
      eprint={2208.06561},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2208.06561}, 
}