RUOD-R: A High-Fidelity Re-Annotated Benchmark for Underwater Object Detection

This repository contains the work from the paper RUOD-R: A High-Fidelity Re-Annotated Benchmark for Underwater Object Detection, introducing revised annotations (RUOD-R) for the Real-world Underwater Object Detection (RUOD) dataset.

Train/test splits, file names, image IDs, and resolutions are unchanged from the original release. RUOD-R provides ~3.5× more instance annotations across the same 14,000 images and 10 classes.

Layout

Annotations/
├── RUOD annotations/          # Original RUOD (reference)
│   ├── instances_train.json
│   └── instances_test.json
├── RUOD-R/
│   ├── COCO/                  # Full re-annotation
│   └── YOLO/                  # YOLO.zip, classes.txt, data.yaml
└── Filtered RUOD-R/           # Subset of images/labels
    ├── COCO/
    └── YOLO/

Figures/                       # Visualizations for this repo
Labelers assignment sheet/     # Annotator assignments & labeling log (xlsx)

Class Distribution Across RUOD Annotation Versions

Figure: Class distribution across the three UOD annotation versions (RUOD, RUOD-R, and Filtered RUOD-R).

Images

This repository provides annotations only; it does not include the RUOD images. Download the image archives from the RUOD dataset repository.

Filtered RUOD-R: For a more balanced label distribution and a more practical training/testing set, we create a filtered version of the revised dataset by excluding images with more than 100 bounding boxes per image. The resulting filtered dataset has only 302 fewer images, but over 70k fewer instances (bounding boxes) compared with RUOD-R.

Formats

• COCO: standard fields; bounding boxes are [x_min, y_min, width, height] in pixels (top-left origin).
• YOLO: one row per object: class_id x_center y_center width height with all coordinates normalized to [0, 1]; class_id is COCO category_id minus 1.

Annotation utilities

If you want to modify the annotations, for example after changing image resolution, or filtering images by bounding-box counts, you can use the annotation utilities in this repository.

Citation

If you use RUOD-R, please cite this paper:

@ARTICLE{11483160,
  author={Awad, Ali and Saleem, Ashraf and Aljnadi, Yaman and Lucas, Evan and Paheding, Sidike and Havens, Timothy C.},
  journal={IEEE Access}, 
  title={RUOD-R: A High-Fidelity Re-Annotated Benchmark for Underwater Object Detection}, 
  year={2026},
  volume={},
  number={},
  pages={1-1},
  keywords={Filtering;Feedback;Filters;Circuits;Location awareness;Protocols;Mobile communication;Communication systems;Pixel;Electronic mail;Underwater object detection;Image enhancement;Dataset re-annotation;Bounding box quality;Label noise;Deep learning;Marine robotics;Benchmark evaluation},
  doi={10.1109/ACCESS.2026.3685121}}

Because RUOD-R builds on the Real-world Underwater Object Detection (RUOD) benchmark, you should also cite the original RUOD dataset paper (Rethinking general underwater object detection: Datasets, challenges, and solutions):

@article{fu2023rethinking,
  title={Rethinking general underwater object detection: Datasets, challenges, and solutions},
  author={Fu, Chenping and Liu, Risheng and Fan, Xin and Chen, Puyang and Fu, Hao and Yuan, Wanqi and Zhu, Ming and Luo, Zhongxuan},
  journal={Neurocomputing},
  volume={517},
  pages={243--256},
  year={2023},
  publisher={Elsevier}
}