SeaDronesSee: Maritime UAV Object Detection Dataset

Unofficial redistribution of the SeaDronesSee object-detection (v2) dataset, reformatted into a standardized YOLO-compatible directory layout.

Disclaimer

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

SeaDronesSee was created by Leon Amadeus Varga, Benjamin Kiefer, Martin Messmer, and Andreas Zell at the University of Tübingen, who retain all copyright and intellectual property rights (to the extent applicable under the dataset's public-domain-style license — see License below). This repository does not claim ownership of any images, annotations, or metadata.

This repository exists for two purposes:

1. To reorganize the dataset into a standardized YOLO/Ultralytics-compatible directory structure that can be used directly by many modern object detection training pipelines.
2. To provide a more reliable download source, as the original hosting may be slow or difficult to access.

This redistribution is sourced directly from the official "CompressedVersion" release (already COCO-annotated), not from a third-party mirror. The original, uncompressed release (~120GB, full 4K imagery) is not mirrored here -- this repository covers the CompressedVersion only.

---

Dataset Description

SeaDronesSee is a maritime search-and-rescue benchmark: images and video captured by UAVs (drones) over open water, annotated for detecting swimmers, boats, jet skis, life-saving appliances, and buoys. It is designed to support drone-based maritime search-and-rescue research.

This repository covers the Object Detection v2 track specifically (there are other tracks in the full SeaDronesSee benchmark — tracking, multi-spectral detection, and a Boat-MNIST classification set — none of which are included here).

No public test-set labels. The official images/test/ split (3,750 images) has no accompanying annotation file — it is a held-out set for the benchmark's evaluation server. This repository therefore only provides train and val splits with labels; evaluate on val, not test.

Images are not uniformly sized. Unlike some other datasets in this collection, SeaDronesSee's imagery spans several resolutions (1920x1080, 3840x2160, and others) depending on the source drone/camera -- there is no single fixed frame size across the dataset.

---

Changes from the Official Release

• Category filtering. The official annotation files declare 6 categories, including ignored (id 0) — a "don't-care" region marker for ambiguous areas, not a real object class. In the specific release this repository is built from, ignored has zero actual annotations, so it is dropped from the exported class list entirely (5 classes, re-indexed from 0) rather than keeping an always-empty 6th class. No real annotations were removed by this filtering.
• Format converted. Original annotations are COCO-JSON (instances_{train,val}.json); this repository additionally provides a YOLO-format export (normalized class x_center y_center width height .txt files), alongside the original-style COCO layout.
• No test split included, for the reason above — this is an omission relative to the full image set (which does include images/test/), not a modification of any existing data.
• Sharded for upload. Each split's images/labels are chunked into fixed-size shard_NNN/ subdirectories (3,000 files per shard) purely as an artifact of how this repository was uploaded -- it has no effect on the data itself. Use data.yaml's train:/val: lists (which already enumerate every shard) to train across the full split.
• No images were modified. No annotated boxes were changed, added, or removed (beyond the ignored-category filtering, which affected zero boxes in practice).

---

Dataset Structure

dataset/
├── README.md
├── data.yaml
├── data/
│   ├── images/
│   │   ├── train/
│   │   │   ├── shard_000/  (includes metadata.jsonl)
│   │   │   ├── shard_001/
│   │   │   └── shard_002/
│   │   └── val/
│   │       └── shard_000/
│   └── labels/
│       ├── train/
│       │   ├── shard_000/
│       │   ├── shard_001/
│       │   └── shard_002/
│       └── val/
│           └── shard_000/

where:

• images/ contains the drone-captured RGB images for each split, chunked into shard_NNN/ subdirectories.
• labels/ contains one YOLO-format .txt annotation file per image (class x_center y_center width height, normalized), in matching shard_NNN/ subdirectories.
• Each image shard directory also contains a metadata.jsonl (one JSON line per image, with absolute-pixel objects.bbox/objects.categories) so the Hugging Face Dataset Viewer can render this dataset natively despite the sharded YOLO layout.
• data.yaml is the Ultralytics dataset configuration file (class names, split paths, already listing every shard). Note: unlike some other datasets in this collection, no test: key is listed at all here, since there is no public test-set to evaluate against.
• Splits: train 8,930 images / 57,760 boxes (3 shards) · val 1,547 images / 9,630 boxes (1 shard) — 10,477 total; matches the official Object Detection v2 track's published train/val counts exactly.

Classes (5)

swimmer, boat, jetski, life_saving_appliances, buoy

---

Dataset Sources

Original Paper

SeaDronesSee: A Maritime Benchmark for Detecting Humans in Open Water

Leon Amadeus Varga, Benjamin Kiefer, Martin Messmer, Andreas Zell

Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022, pages 2260-2270.

Official Resources

• Official Repository: https://github.com/Ben93kie/SeaDronesSee
• Official Website: https://seadronessee.cs.uni-tuebingen.de/

---

Attribution

All credit for the dataset belongs entirely to the original authors: Leon Amadeus Varga, Benjamin Kiefer, Martin Messmer, and Andreas Zell (University of Tübingen).

This repository only reorganizes their data into a YOLO-compatible layout, for improved usability and accessibility.

If you use this dataset in your research, please cite the original publication below, even though the license does not legally require it (see below).

---

License

Per the official repository, the SeaDronesSee dataset is released under Creative Commons Zero 1.0 Universal (CC0 1.0) — a public-domain dedication placing the data in the public domain with no usage restrictions and no attribution requirement (the accompanying code in the official repository is separately MIT-licensed, which does not apply to this data-only redistribution).

Accordingly:

• No attribution is legally required, though it is appreciated and this card asks that you cite the original paper as a courtesy.
• Commercial and non-commercial use are both permitted.
• No share-alike obligation.

This repository is distributed under the same CC0 1.0 dedication.

---

Citation

If you use this dataset, please consider citing:

@inproceedings{varga2022seadronessee,
  title={SeaDronesSee: A maritime benchmark for detecting humans in open water},
  author={Varga, Leon Amadeus and Kiefer, Benjamin and Messmer, Martin and Zell, Andreas},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={2260--2270},
  year={2022}
}

@misc{varga2021seadronesseemaritimebenchmarkdetecting,
      title={SeaDronesSee: A Maritime Benchmark for Detecting Humans in Open Water},
      author={Leon Amadeus Varga and Benjamin Kiefer and Martin Messmer and Andreas Zell},
      year={2021},
      eprint={2105.01922},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2105.01922},
}

---

Acknowledgements

We sincerely thank Leon Amadeus Varga, Benjamin Kiefer, Martin Messmer, and Andreas Zell for creating and publicly releasing this valuable maritime search-and-rescue benchmark.