---
license: cc-by-nc-4.0
pretty_name: KoSim-GL
tags:
  - uav
  - geo-localization
  - cross-view-retrieval
  - simulation
---

# KoSim-GL

KoSim-GL is a benchmark for matching simulated UAV views to satellite imagery
in urban areas of Daejeon, Republic of Korea. The UAV images were generated
with an AirSim and ROS flight simulator. The satellite imagery was collected
through the Google Maps Tile API. See the [project repository](https://github.com/ahn-heejin/KoSim-GL)
and [paper](https://doi.org/10.3390/electronics15122720) for the collection
method and benchmark protocol.

## Dataset contents

| Split | Locations | UAV images | Satellite images |
| --- | ---: | ---: | ---: |
| Train | 255 | 2,188,425 | 1,530 |
| Test | 29 | 261,890 | 174 |
| Train + test | 284 | 2,450,315 | 1,704 |

The archive also contains a `subset_eval` directory with 162 query UAV images
and 162 gallery satellite images. These 324 additional image entries are not
included in the train and test totals above. The archive also contains six
text metadata files and one Python counting script.

The full dataset is a 394,614,137,420-byte ZIP, distributed as ten consecutive
binary parts. It contains the train and test splits in a
University-1652-compatible layout. UAV views cover six altitude levels from
100 to 600 m and five camera views (one nadir and four oblique).

## Download and reconstruction

Download all ten files named `20261013_multi_sinsungdong_data_v2.zip.part01`
through `20261013_multi_sinsungdong_data_v2.zip.part10` from
[Files and versions](https://huggingface.co/datasets/hhhuugging/KoSim-GL/tree/main)
into one directory. Each part is 39,461,413,742 bytes. Use only the `v2`
parts and keep their names unchanged. Run this Python code from that directory
to restore the original ZIP:

```python
from pathlib import Path
import hashlib

base = "20261013_multi_sinsungdong_data_v2.zip"
digest = hashlib.sha256()
with Path("multi_sinsungdong_data.zip").open("xb") as output:
    for number in range(1, 11):
        with Path(f"{base}.part{number:02d}").open("rb") as source:
            while block := source.read(16 * 1024 * 1024):
                output.write(block)
                digest.update(block)

assert digest.hexdigest() == "c6a51312f03f001210cb58e675d050514bfbe596890ff6f66f1f128194d85840"
```

The SHA-256 above is for the original unsplit ZIP. Individual part checksums
are in the [machine-readable manifest](https://huggingface.co/datasets/hhhuugging/KoSim-GL/blob/main/20261013_KoSim-GL_virtual_parts_manifest_v3.json).
Reconstruction needs approximately 395 GB of free space in addition to the
downloaded parts.

## Intended use

Research and evaluation of cross-view image retrieval and UAV
geo-localization. The [project repository](https://github.com/ahn-heejin/KoSim-GL)
lists the dataset license as Creative Commons Attribution-NonCommercial 4.0.

## Citation

Ahn, H., Lee, C., Lee, S., Wi, H., Jang, I., and Choi, D.-G.
"KoSim-GL: A Large-Scale Simulation-Based Dataset for UAV Cross-View
Geo-Localization in Korean Urban Environments." *Electronics* 15(12), 2720
(2026). [https://doi.org/10.3390/electronics15122720](https://doi.org/10.3390/electronics15122720)

## Contact

For dataset questions, open an issue in the
[project repository](https://github.com/ahn-heejin/KoSim-GL/issues).