AerialVL Dataset

This repo covers the supplementary materials of the following paper:

AerialVL: A Dataset, Baseline and Algorithm Framework for Aerial-Based Visual Localization With Reference Map

Dataset

Download

Two parts of the dataset can be downloaded from the Tsinghua Cloud or Baidu Netdisk, including the captured frames and satellite imageries for training.

Description

• Flight data covering about 70 km trajectories with various terrains, multiple heights and illumination changes.
• For sequence-based visual localization, 11 frame sequences with different paths ranging from the shortest one of 3.7 km to the longest up to 11 km.
• For visual place recognition, 18361 separate aerial-based images with 14096 cropped corresponding map patches are provided.

Sensor Setup

RGB camera (FLIR BFS-U3-31S4C-C) is attached with a gimbal to the UAV. The GNSS (NovAtel OEM718D) has 1.5 m accuracy in RMS with the single point mode.

Evaluation

VAL

In the sequence-based visual localization part, we have prepared the evaluation data as follows.

+--- geo_referenced_map
|   +--- @large_map@120.42114259488751@36.604504047017464@120.48431398161503@36.573629616877625@.tif
|   +--- @small_map@120.42114259488751@36.604504047017464@120.4568481612987@36.586863027841225@.tif
+--- long_trajtr
|   +--- 2023-03-16-18-04-01
|   +--- 2023-03-18-12-18-25
|   +--- 2023-03-18-12-47-05
|   +--- 2023-03-18-14-38-32
|   +--- 2023-03-18-15-01-14
|   +--- 2023-03-18-15-40-18
+--- short_trajtr
|   +--- 2023-03-11-11-48-35
|   +--- 2023-03-16-16-58-43
|   +--- 2023-03-18-16-30-27
|   +--- 2023-03-18-16-43-16
|   +--- 2023-03-18-16-55-37

There are two geo-referenced map with different size for flight sequences with different length. They are also renamed as the the following formats (maps are heading north):

@map_name@LeftTopLongitude@LeftTopLatitude@RightBottomLongitude@RightBottomLatitude@.tif

And the captured frames are also re-organized as @UTCTimeStamp@Longitude@Latitude@.png (frames are heading east).

VPR

The visual place recognition part are presented as follows:

+--- map_database
|   +--- level_1
|   +--- level_2
|   +--- level_3
+--- query_images
|   +--- query_images_1
|   +--- query_images_2
|   +--- query_images_3
|   +--- query_images_4
+--- raw_satellite_imagery
|   +--- @map@120.42251588590332@36.60395282621937@120.48225404509132@36.573629616877625@.tif

The map tiles in the map_database folder are sampled from the satellite imagery, which is downloaded from the Google Earth. The different levels in the map_database present the tiles with different size. These tiles are re-organized as follows (tiles are heading east to be consistent with the captured frame):

@map@LeftBottomLongitude@LeftBottomLatitude@RightTopLongitude@RightTopLatitude@.png

It is worth mentioning that the definition here is different from the VAL part because we adjust the heading of these tiles to make the VPR task more easier.

The query_imagefolder contains the capture four parts of captured frames with names as @Longitude@Latitude@.png.

Training Data

We have also provided a lot of satellite imageries collected from different years using USGS. As the training data, these imageries can help you to get a new VPR model designed for the aerial-based platform.

Citation

If you find this dataset useful for your research, please consider citing the paper

@article{he2024aerialvl,
  author={He, Mengfan and Chen, Chao and Liu, Jiacheng and Li, Chunyu and Lyu, Xu and Huang, Guoquan and Meng, Ziyang},
  journal={IEEE Robotics and Automation Letters}, 
  title={AerialVL: A Dataset, Baseline and Algorithm Framework for Aerial-Based Visual Localization With Reference Map}, 
  year={2024},
  volume={9},
  number={10},
  pages={8210-8217},
  publisher={IEEE}
}