HRSID: High-Resolution SAR Images Dataset (Ship Detection)

Unofficial redistribution of the HRSID high-resolution SAR ship-detection dataset, reformatted into a standardized YOLO-compatible directory layout. License status is unclear -- see License before using this beyond research.

Dataset Description

• Homepage: https://github.com/chaozhong2010/HRSID
• Repository: https://github.com/chaozhong2010/HRSID
• Paper: https://doi.org/10.1109/ACCESS.2020.3005861
• Point of Contact: chaozhong2010@163.com

Disclaimer

This repository is not an official release of HRSID.

HRSID was created by Shunjun Wei, Xiangfeng Zeng, Qizhe Qu, Mou Wang, Hao Su, and Jun Shi and released via the official GitHub repository, who retain all copyright and intellectual property rights. This repository does not claim ownership of any images or annotations.

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 train/validation/test split (the official release ships train/test only) suitable for standard training workflows.

License is genuinely unclear -- read this before using the data for anything beyond research. See License below for the full explanation; in short, the source repository carries a software license (GPL-3.0) whose applicability to the dataset itself is not stated, and part of the source imagery may be subject to additional restrictions from its originating satellite programs. This card is deliberately explicit about that uncertainty rather than picking a license label that cannot be verified.

---

Dataset Overview

HRSID is a high-resolution Synthetic Aperture Radar (SAR) benchmark for ship detection, semantic segmentation, and instance segmentation, modeled on COCO's construction process. It contains 5,604 800x800 SAR image crops (drawn from 136 larger scenes) with 16,951 ship instances, spanning multiple resolutions (0.5m, 1m, 3m), polarizations, sea states, and coastal/open-sea conditions. Source imagery is drawn from Sentinel-1B, TerraSAR-X, and TanDEM-X satellite SAR data.

This repository preserves every image and bounding box while re-encoding the labels for YOLO compatibility and reorganizing the official split (see Changes from the Official Release below).

---

Changes from the Official Release

• Format converted. The official release ships COCO-format annotations (train2017.json / test2017.json, polygon segmentation + bounding boxes). This repository re-encodes the bounding boxes as YOLO's normalized class x_center y_center width height .txt format (one line per box) and a matching canonical COCO JSON; polygon segmentation is dropped (this card targets bounding-box detection, not instance segmentation).
• A validation split was carved out of train2017. The official release has no validation set. This repository holds out a seeded random 15% of train2017's 3,642 crops as valid (546 images), leaving 3,096 for train; test2017 (1,962 images) is kept as test, unchanged. The official split is itself a random split over crops (not scenes), so this follows the same methodology.
• Single class remapped. HRSID's one category (ship, source id 1) is remapped to id 0 for a contiguous 0-indexed class list.
• No image pixel content was modified. No boxes were added or removed.

---

Dataset Structure

<repo>/
├── README.md
├── hrsid_banner.jpg
└── data/
    ├── data.yaml
    ├── images/
    │   ├── train/   (3,096 *.jpg)
    │   ├── valid/   (546 *.jpg)
    │   └── test/    (1,962 *.jpg)
    └── labels/
        ├── train/   (3,096 *.txt)
        ├── valid/
        └── test/

where:

• data/images/<split>/ contains the 800x800 SAR image crops for each split.
• data/labels/<split>/ contains one YOLO-format .txt annotation file per image (class x_center y_center width height, normalized).
• data/data.yaml is the Ultralytics dataset configuration file (class names, split paths, relative to data/).
• Splits: train 3,096 images / 9,143 boxes · valid 546 images / 1,904 boxes · test 1,962 images / 5,922 boxes (5,604 images / 16,969 boxes total).

Classes (1)

ship -- this is a single-class dataset; images vary from zero to many annotated ships per crop.

---

Dataset Sources

Original Paper

HRSID: A High-Resolution SAR Images Dataset for Ship Detection and Instance Segmentation

Shunjun Wei, Xiangfeng Zeng, Qizhe Qu, Mou Wang, Hao Su, Jun Shi

IEEE Access, vol. 8, pp. 120234-120254, 2020. DOI: 10.1109/ACCESS.2020.3005861

Official Resources

• Official Repository: https://github.com/chaozhong2010/HRSID
• Contact: chaozhong2010@163.com

---

Attribution

All credit for the dataset belongs entirely to the original HRSID authors: Shunjun Wei, Xiangfeng Zeng, Qizhe Qu, Mou Wang, Hao Su, and Jun Shi.

This repository only reorganizes their data into a YOLO-compatible layout with a train/valid/test split, for improved usability.

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

---

License

Marked unknown on this repository because it genuinely cannot be verified from public information -- read this section before relying on it.

• The official GitHub repository contains a LICENSE file that is the GNU General Public License v3.0 (GPL-3.0), a software license. Nothing in the repository's README or data files states whether this is intended to cover the dataset itself (as opposed to any accompanying code), and the dataset's own COCO annotation files carry an empty licenses field.
• HRSID's source imagery is drawn from Sentinel-1B (ESA/Copernicus, open data) and TerraSAR-X / TanDEM-X (DLR -- distributed under a scientific-use license that does not, in general, permit free redistribution of derived image products). It is not possible to determine from public information which crops in this dataset derive from which sensor, or whether the authors obtained redistribution rights for the TerraSAR-X/TanDEM-X-derived portion.
• Given that, this repository makes no license claim and is tagged unknown rather than an SPDX identifier that cannot be substantiated.

If you plan to use this dataset for anything beyond personal research experimentation (redistribution, commercial use, a downstream dataset release), we recommend contacting the original authors directly to confirm terms, particularly regarding the TerraSAR-X/TanDEM-X-derived imagery.

---

Citation

If you use this dataset, please cite:

@ARTICLE{wei2020hrsid,
  author={Wei, Shunjun and Zeng, Xiangfeng and Qu, Qizhe and Wang, Mou and Su, Hao and Shi, Jun},
  journal={IEEE Access},
  title={HRSID: A High-Resolution SAR Images Dataset for Ship Detection and Instance Segmentation},
  year={2020},
  volume={8},
  pages={120234-120254},
  doi={10.1109/ACCESS.2020.3005861}
}

---

Acknowledgements

We sincerely thank Shunjun Wei, Xiangfeng Zeng, Qizhe Qu, Mou Wang, Hao Su, and Jun Shi for creating and publicly releasing this valuable SAR ship-detection benchmark.