Aeroscapes: Aerial Semantic Segmentation Dataset

Unofficial redistribution of the AeroScapes dataset under the original CC BY-SA 4.0 license.

Disclaimer

This repository is not an official release of the AeroScapes dataset.

The AeroScapes dataset was created by Ishan Nigam, Chen Huang, and Deva Ramanan, with dataset collection and manual annotation supported by Autel Robotics. They retain all copyright and intellectual property rights. This repository does not claim ownership of any images, annotations, or metadata.

This repository exists for two purposes:

1. To reorganize the original dataset into a standardized YOLO/Ultralytics-compatible directory structure (also consumable by CABiNet) that can be used directly by many modern semantic segmentation training pipelines.
2. To provide a more convenient download source, as the original is distributed as a single Google Drive archive.

---

Dataset Description

AeroScapes is a semantic segmentation benchmark captured from low-altitude unmanned aerial vehicles (UAVs), acquired at altitudes between 5 and 50 meters. The dataset contains 3,269 720p (1280x720) images with pixel-level semantic annotations across 12 classes (11 foreground classes plus background), covering people, vehicles, and both urban and natural terrain.

This repository preserves the original dataset while packaging it in a standardized directory layout for improved compatibility with modern deep learning frameworks.

---

Changes from the Official Release

This repository does not modify the dataset contents.

The following changes have been made:

• Reorganized the directory structure into a YOLO/Ultralytics-compatible images/+masks/ layout.
• Added a data.yaml configuration file for easier integration with Ultralytics-based projects.
• Preserved the original train and validation splits exactly as released (ImageSets/trn.txt and val.txt) — the official AeroScapes release does not define a separate test split.
• Preserved all original filenames.
• Preserved all original images.
• Preserved all original segmentation masks — the official SegmentationClass/ masks are already single-channel, class-ID-encoded PNGs, so this is a lossless copy/re-layout, not a re-annotation or re-encoding.
• No labels were changed.
• No samples were added or removed.

Apart from the directory organization and configuration file, the dataset contents are identical to the official release.

---

Dataset Structure

dataset/
├── README.md
├── data.yaml
├── images/
│   ├── train/    # 2,621 images
│   └── val/      # 648 images
└── masks/
    ├── train/
    └── val/

where:

• images/ contains the original RGB UAV images organized by dataset split.
• masks/ contains the corresponding single-channel semantic segmentation masks (pixel value = class ID, 0-11) for each split.
• data.yaml is a configuration file added in this repository to simplify loading the dataset in Ultralytics-compatible training pipelines.
• The original train/validation splits have been preserved exactly as released by the AeroScapes authors. There is no official test split.

Each image in images/<split>/ has a corresponding segmentation mask with the same filename (different extension) in masks/<split>/.

Classes

ID	Class	ID	Class
0	Background	6	Animal
1	Person	7	Obstacle
2	Bike	8	Construction
3	Car	9	Vegetation
4	Drone	10	Road
5	Boat	11	Sky

All 12 classes are valid and used in both training and evaluation; none are mapped to an ignore label. Pixel value 255 is reserved for genuinely unrecognized values (none are expected in a clean copy of this dataset).

---

Dataset Sources

Original Paper

Ensemble Knowledge Transfer for Semantic Segmentation

Ishan Nigam, Chen Huang, Deva Ramanan

2018 IEEE Winter Conference on Applications of Computer Vision (WACV)

---

Official Resources

• Official Repository: https://github.com/ishann/aeroscapes
• Dataset Download (Google Drive): https://drive.google.com/file/d/1WmXcm0IamIA0QPpyxRfWKnicxZByA60v/view?usp=sharing

---

Attribution

All credit for the dataset belongs entirely to the original AeroScapes authors and Autel Robotics (dataset collection and manual annotation support).

This repository only redistributes the original dataset under the same license while reorganizing the directory structure for improved usability and accessibility.

If you use this dataset in your research, please cite the original publication below.

---

License

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

Accordingly:

• Attribution to the original authors is required.
• Commercial use is permitted.
• Any derivative work must be distributed under the same license.

This repository is distributed under the same license.

---

Citation

If you use this dataset, please cite:

@inproceedings{nigam2018ensemble,
  author    = {Ishan Nigam and Chen Huang and Deva Ramanan},
  title     = {Ensemble Knowledge Transfer for Semantic Segmentation},
  booktitle = {2018 IEEE Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2018}
}

---

Acknowledgements

We sincerely thank Ishan Nigam, Chen Huang, and Deva Ramanan for creating and publicly releasing this valuable benchmark, and Autel Robotics for supporting its collection and annotation — this work has significantly contributed to research in semantic segmentation for UAV imagery.