osmarluiz/BSB-Aerial-Dataset

---
license: mit
task_categories:
- image-segmentation
- object-detection
tags:
- panoptic-segmentation
- aerial-imagery
- remote-sensing
- COCO
- detectron2
- urban
- Brasilia
pretty_name: BSB Aerial Dataset
size_categories:
- 1K<n<10K
language:
- en
---

# BSB Aerial Dataset

A panoptic segmentation dataset of aerial imagery from Brasilia, Brazil, annotated in COCO format.

## Dataset Description

The BSB Aerial Dataset contains **3,400 aerial image tiles** (512x512 pixels) of urban areas in Brasilia, annotated for panoptic segmentation with **14 categories** covering both "stuff" and "things" classes.

### Splits

| Split | Images |
|-------|--------|
| Train | 3,000 |
| Val | 200 |
| Test | 200 |

### Categories

| ID | Name | Type |
|----|-----------------------|-------|
| 1 | Street | Stuff |
| 2 | Permeable Area | Stuff |
| 3 | Lake | Stuff |
| 4 | Swimming Pool | Thing |
| 5 | Harbor | Thing |
| 6 | Vehicle | Thing |
| 7 | Boat | Thing |
| 8 | Sports Court | Thing |
| 9 | Soccer Field | Thing |
| 10 | Comm. Building | Thing |
| 11 | Comm. Building Block | Thing |
| 12 | Res. Building | Thing |
| 13 | House | Thing |
| 14 | Small Construction | Thing |

## Dataset Structure

```
bsb_dataset/
├── annotations/
│ ├── panoptic_train.json
│ ├── panoptic_val.json
│ ├── panoptic_test.json
│ ├── instance_train.json
│ ├── instance_val.json
│ └── instance_test.json
├── image_train/ # RGB aerial tiles (TIFF)
├── image_val/
├── image_test/
├── panoptic_train/ # Panoptic segmentation masks
├── panoptic_val/
├── panoptic_test/
├── panoptic_stuff_train/ # Stuff-only masks
├── panoptic_stuff_val/
├── panoptic_stuff_test/
├── class_train/ # Semantic class masks
├── class_val/
└── class_test/
```

## Annotation Format

Annotations follow the [COCO Panoptic format](https://cocodataset.org/#format-data). Each annotation JSON contains image metadata and segment information compatible with tools like Detectron2.

## Usage with Detectron2

See the example notebook in the [GitHub repository](https://github.com/osmarluiz/BSB-Aerial-Dataset) for a full implementation using Detectron2's Panoptic-FPN.

> **Note:** The repository includes a modified `detection_utils.py` that properly handles RGB image tiles. Replace the original Detectron2 file with the provided version.

## Related Tools

- [Panoptic-Generator](https://github.com/abilius-app/Panoptic-Generator) — Tool for building remote sensing panoptic segmentation datasets in COCO format using GIS software.

## Citation

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

```bibtex
@misc{bsb_aerial_dataset,
author = {Osmar Luiz Carvalho},
title = {BSB Aerial Dataset: Panoptic Segmentation of Aerial Imagery from Brasilia},
year = {2024},
url = {https://github.com/osmarluiz/BSB-Aerial-Dataset}
}
```

## License

This dataset is released under the [MIT License](https://opensource.org/licenses/MIT).