AgEagleVPRDataset

overview

The datasets provided by the AgEagle are suitable for the aerial-based VPR task. We collect the corresponding satellite map and make them in a standard format for the VPR task evaluation.

The original datasets are publicly downloaded from the AgEagle website at this link.

Currently, there are several datasets suitable for the aerial-based VPR task, as illustrated below:

Name	Num.Qr	Num.Db	Charac.	Status
University_campus	443	2720	Various Orientation	Completed
Industrial_estate	277	775	Thermal Image	Completed
Village_switzerland	297	1845	Farmland Scene	Completed

Notably, the database is sampled with overlapping and suitable size.

how to use

The structure of provided dataset is the same as the other aerial-based geo-localization benchmark. The dataset is organized in a directory tree as such:

.
└── AgEagleVPR
    └── SubScene-One
        ├── map_database
        │   ├── @SubScene-One@longitude@latitude@patch_size.tif
        │   ├── .......    
        │   └── @SubScene-One@longitude@latitude@patch_size.tif
        │
        └── query_images
            ├── @query_index@longitude@latitude@.png
            ├── .......    
            └── @query_index@longitude@latitude@.png

The retrieval positive radius is set to 200 meters and the distance can be calculated by the coordinates of the image.

baseline

We also simply tested the performance of the SOTA method on this dataset as the baseline. The selected method is SALAD and we retrained this model with a self-collected training dataset.

Under the setting of certain positive radius (following the names of the datasets), the retrieval recalls are given in the table below:

Name	Method	R@1	R@5	R@10
University_campus(200m)	SALAD(retrained)	45.82	67.04	75.85
Industrial_estate(100m)	SALAD(retrained)	74.01	90.25	93.14
Village_switzerland(100m)	SALAD(retrained)	75.42	94.95	97.98

acknowledgements

Part of the AgEagle dataset is used in our previous work. And if you find this dataset useful for your research, please consider citing the paper:

@article{he2024leveraging,
  title={Leveraging map retrieval and alignment for robust UAV visual geo-localization},
  author={He, Mengfan and Liu, Jiacheng and Gu, Pengfei and Meng, Ziyang},
  journal={IEEE Transactions on Instrumentation and Measurement},
  volume={73},
  pages={1--13},
  year={2024},
  publisher={IEEE}
}