---
pretty_name: MARS20
license: afl-3.0
task_categories:
- object-detection
- image-to-text
language:
- en
tags:
- remote-sensing
- aerial-imagery
- airport
- aircraft
- fine-grained-recognition
- keypoint-detection
- image-captioning
- generative-data
annotations_creators:
- expert-generated
language_creators:
- expert-generated
multilinguality:
- monolingual
size_categories:
- 1K<n<10K
source_datasets:
- original
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.parquet
  - split: validation
    path: data/validation.parquet
  - split: test
    path: data/test.parquet
---

# MARS20

MARS20 is a remote-sensing airport dataset for fine-grained aircraft understanding and controllable image generation. This Hugging Face release packages the official split subset as `Parquet` files with embedded RGB images, instance annotations, skeleton keypoints, and image-level captions.

It accompanies the paper **AirplaneGen: Skeleton-Guided Generation of Remote Sensing Images with Multi-Instance Airplanes**.

- Authors: Lingxuan Zhu, Yanze Ma, Jiaji Wu, Yanbo Fan, Xiaobing Wang, Mingzhou Tan
- Venue: *Remote Sensing*, 2026, 18(12):1940
- DOI: <https://doi.org/10.3390/rs18121940>
- Paper: <https://www.mdpi.com/2072-4292/18/12/1940>

## At a Glance

- 2778 annotated images
- 16673 airplane instances
- 20 fine-grained aircraft categories
- Bounding boxes in pixel coordinates
- Skeleton keypoints in normalized coordinates
- Image-level captions

This release includes only the officially split annotated subset. The original local workspace contains 32 extra RGB images without matching official annotations or split membership; they are listed in `metadata/unused_images.json`.

## Splits

| Split | Images | Objects |
|---|---:|---:|
| train | 2563 | 15431 |
| validation | 107 | 566 |
| test | 108 | 676 |

## Categories

`SU-35`, `C-130`, `C-17`, `C-5`, `F-16`, `TU160`, `E-3`, `B-52`, `P-3C`, `B-1B`, `E-8`, `TU-22`, `F-15`, `KC-135`, `F-22`, `FA-18`, `TU-95`, `KC-10`, `SU-34`, `SU-24`

## Schema

Each example contains:

- `id`: image identifier
- `split`: dataset split
- `image`: RGB image
- `width`, `height`: image size
- `image_caption`: image-level caption
- `background_caption`: optional source background text
- `num_objects`: number of airplane instances
- `plane_types`: list of instance classes
- `objects`: full instance annotations

Each item in `objects` contains:

- `plane_type`
- `bbox`: `{xmin, ymin, xmax, ymax}`
- `keypoints`: list of `{label, x, y}`

## Notes

- Most instances contain 8 skeleton keypoints.
- 6 legacy instances contain 6 or 7 keypoints; see `metadata/keypoint_anomalies.json`.
- Caption sources are mixed: `DetailCaption`, `caption-multi.json`, `Caption-Background`, and a small auto-generated subset.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("your-username/MARS20")
sample = ds["train"][0]

print(sample["id"])
print(sample["plane_types"])
print(sample["objects"][0])
```

## Files

- `data/*.parquet`: train/validation/test splits with embedded images
- `metadata/summary.json`: split stats and class counts
- `metadata/class_names.json`: category names
- `metadata/unused_images.json`: excluded source RGB images

## Citation

```bibtex
@article{zhu2026airplanegen,
  author  = {Zhu, Lingxuan and Ma, Yanze and Wu, Jiaji and Fan, Yanbo and Wang, Xiaobing and Tan, Mingzhou},
  title   = {AirplaneGen: Skeleton-Guided Generation of Remote Sensing Images with Multi-Instance Airplanes},
  journal = {Remote Sensing},
  year    = {2026},
  volume  = {18},
  number  = {12},
  pages   = {1940},
  doi     = {10.3390/rs18121940},
  url     = {https://doi.org/10.3390/rs18121940}
}
```