KFGOD (KOMPSAT Fine-Grained Object Detection Dataset)

1) Overview

KFGOD is a large-scale fine-grained object detection dataset built from high-resolution KOMPSAT-3 / KOMPSAT-3A satellite imagery (approximately 0.55–0.7m GSD).
It provides 33 fine-grained classes and around 882,399 annotated object instances, with annotations available as both:

• OBB (Oriented Bounding Box)
• HBB (Horizontal Bounding Box)

KFGOD was designed as a homogeneous benchmark dataset by using KOMPSAT imagery only, reducing the sensor domain gap and enabling more consistent evaluation across methods.

---

2) Dataset Splits

This Hugging Face release provides the following splits:

• train: public
• validation: public
• test: hidden (for benchmark evaluation)

Note: The test split is intentionally not public due to fair benchmarking purposes.

According to the paper, the split sizes are:

• Train: 3073 images
• Validation: 470 images
• Test: 460 images

---

3) Classes (Labels)

KFGOD contains 33 fine-grained classes, grouped into 5 parent categories:

• Ship (10)
• Aircraft (5)
• Vehicle (4)
• Container (2)
• Infrastructure (12)

Ship (10)

• motorboat (MB), sailboat (SB), tugboat (TB), barge (BG), fishing boat (FB)
• ferry (FR), container ship (CS), oil tanker (OT), drill ship (DS), warship (WS)

Aircraft (5)

• fighter aircraft (FA), large military aircraft (LM)
• small civilian aircraft (SC), large civilian aircraft (LC), helicopter (HC)

Vehicle (4)

• small vehicle (SV), truck (TR), bus (BS), train (TN)

Container (2)

• container (CT), container group (CG)

Infrastructure (12)

• crane (CR), bridge (BR), dam (DM), storage tank (ST), sports field (SF)
• stadium (SD), swimming pool (SP), roundabout (RA), helipad (HP), wind turbine (WT)
• aquaculture farm (AF), marine research station (MR)

---

4) Annotation Format

KFGOD annotations were originally labeled using OBB (Oriented Bounding Boxes), and can be converted into multiple popular formats for broader training/evaluation compatibility.

Commonly used formats include:

• OBB / HBB
• COCO / DOTA / YOLO
• GeoJSON

---

5) License

This dataset is distributed under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.

You are allowed to:

• Modify / adapt the dataset
• Use it commercially

You must:

• Provide appropriate attribution to the original dataset and paper
• Share derived versions under the same license (CC BY-SA 4.0)

Full license text: https://creativecommons.org/licenses/by-sa/4.0/

---

6) Paper

Please refer to the official paper for full details:

KFGOD: A Fine-Grained Object Detection Dataset in KOMPSAT Satellite Imagery
Remote Sensing 2025, 17, 3774
DOI: https://doi.org/10.3390/rs17223774

---

7) Original Dataset Source (KISTI DataON)

The original KFGOD dataset is hosted on KISTI Korea Research Data Platform (DataON).
This Hugging Face release is provided as a convenient distribution mirror to improve accessibility and usability for the research/developer community.

• Original DOI (DataON): https://doi.org/10.22711/idr/1101

---

8) Citation

If you use this dataset in your research, publications, or products, please cite:

@article{lee2025kfgod,
  title={KFGOD: A Fine-Grained Object Detection Dataset in KOMPSAT Satellite Imagery},
  author={Lee, Dong Ho and Hong, Ji Hun and Seo, Hyun Woo and Oh, Han},
  journal={Remote Sensing},
  volume={17},
  year={2025},
  pages={3774},
  doi={10.3390/rs17223774}
}