osmarluiz/BSB-Aerial-Dataset

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. 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 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 — 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:

@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.