AQUABENCH: Evaluating Vision Foundation Models for Underwater Segmentation

AQUABENCH is a unified semantic segmentation benchmark for evaluating how well pretrained vision foundation models (VFMs) transfer to underwater imagery. It repurposes nine publicly available underwater datasets into a common image–mask interface with standardized splits, resolutions, and background conventions, while preserving each dataset's original class taxonomy.

• Paper: AQUABENCH: Evaluating Vision Foundation Models for Underwater Segmentation, T. Globisch and S. Oehmcke, ECCV 2026 Workshops (Marine Vision)
• Code (conversion pipeline + benchmark interface, MIT): <link to repository>
• Contact: torben.globisch@uni-rostock.de, stefan.oehmcke@uni-rostock.de (University of Rostock)

Subsets

Each dataset is a separate config. #Cls includes the background class (index 0).

Config	Dataset	Domain	Train	Val	Test	#Cls	Size (H×W)	License
coralscapes	Coralscapes	Coral reef health	1,517	166	392	39	512×1024	Apache-2.0
cou	Common Objects Underwater (COU)	Man-made objects, pool/lake/ocean	6,753	1,952	958	24	576×1024	CC BY-SA 4.0
deepfish	DeepFish	Fish habitats	310	124	186	2	576×1024	MIT
l4s	Looking for Seagrass (L4S)	Seagrass coverage	4,223	610	1,204	2	576×1024	BSD-2-Clause
liaci	LIACi	Ship hull inspection (ROV)	1,233	137	191	11	576×1024	CC BY-NC-SA 4.0
seaclear	SeaClear	Marine debris, shallow water	6,071	674	1,865	41	576×1024	CC BY 4.0
suim	SUIM	Scene-level underwater	1,297	228	110	8	480×640	MIT
trashcan	TrashCan 1.0	Deep-sea debris	5,459	606	1,147	23	576×1024	Non-commercial (JAMSTEC)
uiis10k	UIIS10K	Aggregated underwater scenes	7,234	804	2,010	11	480×640	Apache-2.0
Total			34,097	5,301	8,063

Usage

from datasets import load_dataset

ds = load_dataset("TorbenGl/AQUABENCH", "suim", split="train")
sample = ds[0]
image, mask = sample["image"], sample["mask"]   # RGB image, class-index mask

Per-config metadata (class vocabulary, ignore indices, split sizes) is in _metadata/<config>.json.

Data format

Every config stores rows of (image, mask) as PNG-compressed bytes in parquet files, one file set per split.

• Masks are single-channel class-index maps. Index 0 = background in all configs. COCO datasets without an explicit background polygon get the background filled automatically.
• Evaluation: mIoU over all foreground classes, excluding background.

Preprocessing

The same deterministic transformation applies to all models.

1. Task conversion. COCO datasets (cou, liaci, seaclear, trashcan, uiis10k) are flattened from RLE/polygons to per-pixel class masks. Where annotations overlap, the later-drawn label wins. For LIACi the draw order is set so that surface conditions (corrosion, paint peeling, defects) are drawn on top of structural classes. For all other COCO sets the raw order is kept, with <5% single-pair overwrite. Image+mask datasets are remapped from grayscale values (l4s, deepfish) or RGB colours (suim).
2. Aspect-ratio normalization. Each dataset is assigned to the family closest to its median aspect ratio: 16:9 → 576×1024, 2:1 → 512×1024 or 4:3 → 480×640. All sizes are multiples of 16, giving exact patch-16 grids.
3. Resize and crop. Each image is resized to cover the target size while keeping its aspect ratio (bilinear for images, nearest-neighbour for masks), so nothing is stretched or padded. It is then cropped along the oversized axis at the position that keeps the most foreground pixels, and no annotated class is removed. Images without foreground are center-cropped.
4. Splits. Official splits are used where available. If only train/val exist, val becomes test. SeaClear has no predefined split, so train and test are partitioned at the dive level to avoid leakage between highly correlated frames. liaci, seaclear, trashcan and uiis10k get a stratified 10% validation holdout carved from train, stratified by class presence with seed 42. The validation split is used only for monitoring, not for model selection.

Excluded variants: the material and supplementary labels of SeaClear, the material variant of TrashCan, the original-label variant of Coralscapes, CoralMask, and USOD10K.

Known overlaps

Duplicates were detected with perceptual hashing (dHash, Hamming distance ≤ 4). UIIS10K shares about 1,533 near-duplicate images with SUIM and about 165 with TrashCan, so the subsets are not fully independent. See the paper appendix for the ranking sensitivity analysis.

License

AQUABENCH does not relicense the data. Each subset keeps its original ownership, copyright and license; see <config>/dataset_license.txt.

• Using the full benchmark means complying with all nine licenses.
• Using a subset means complying only with the licenses of that subset.
• Non-commercial: liaci (CC BY-NC-SA 4.0) and trashcan (academic and personal use only). Commercial use of TrashCan requires prior permission from JAMSTEC, and attribution to JAMSTEC J-EDI is required in all cases.
• Share-alike: derivatives of cou (CC BY-SA 4.0) and liaci (CC BY-NC-SA 4.0) must be released under the same license.
• uiis10k contains images that originate from SUIM and TrashCan. The terms of those upstream sources may also apply to them.
• The AQUABENCH conversion pipeline and benchmark interface are released under the MIT License, which does not cover the data.

All rights in the constituent datasets remain with their respective rights holders.

Citation

If you use AQUABENCH, cite the benchmark and the original publication of every subset you use.

@inproceedings{globisch2026aquabench,
  title     = {{AQUABENCH}: Evaluating Vision Foundation Models for Underwater Segmentation},
  author    = {Globisch, Torben and Oehmcke, Stefan},
  booktitle = {European Conference on Computer Vision (ECCV) Workshops},
  year      = {2026},
  note      = {To appear}
}

@inproceedings{sauder2025coralscapes,
  title     = {The Coralscapes Dataset: Semantic Scene Understanding in Coral Reefs},
  author    = {Sauder, Jonathan and Domazetoski, Viktor and Banc-Prandi, Guilhem and Perna, Gabriela and Meibom, Anders and Tuia, Devis},
  booktitle = {ICCV},
  pages     = {2115--2122},
  year      = {2025}
}

@inproceedings{mukherjee2025cou,
  title     = {The Common Objects Underwater ({COU}) Dataset for Robust Underwater Object Detection},
  author    = {Mukherjee, Rishi and Singh, Sakshi and McWilliams, Jack and Sattar, Junaed},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages     = {18597--18603},
  year      = {2025}
}

@article{saleh2020deepfish,
  title   = {A realistic fish-habitat dataset to evaluate algorithms for underwater visual analysis},
  author  = {Saleh, Alzayat and Laradji, Issam H. and Konovalov, Dmitry A. and Bradley, Michael and Vazquez, David and Sheaves, Marcus},
  journal = {Scientific Reports},
  volume  = {10},
  number  = {1},
  pages   = {14671},
  year    = {2020},
  doi     = {10.1038/s41598-020-71639-x}
}

@inproceedings{reus2018seagrass,
  title     = {Looking for Seagrass: Deep Learning for Visual Coverage Estimation},
  author    = {Reus, Gereon and M{\"o}ller, Thomas and J{\"a}ger, Jonas and Schultz, Stewart T. and Kruschel, Claudia and Hasenauer, Julian and Wolff, Viviane and Fricke-Neuderth, Klaus},
  booktitle = {2018 OCEANS - MTS/IEEE Kobe Techno-Oceans (OTO)},
  pages     = {1--6},
  year      = {2018},
  doi       = {10.1109/OCEANSKOBE.2018.8559302}
}

@article{waszak2023liaci,
  title   = {Semantic Segmentation in Underwater Ship Inspections: Benchmark and Data Set},
  author  = {Waszak, Maryna and Cardaillac, Alexandre and Elves{\ae}ter, Brian and R{\o}d{\o}len, Frode and Ludvigsen, Martin},
  journal = {IEEE Journal of Oceanic Engineering},
  volume  = {48},
  number  = {2},
  pages   = {462--473},
  year    = {2023},
  doi     = {10.1109/JOE.2022.3219129}
}

@article{duras2024seaclear,
  title   = {A dataset for detection and segmentation of underwater marine debris in shallow waters},
  author  = {{\DJ}ura{\v{s}}, An{\dj}ela and Wolf, Ben J. and Ilioudi, Athanasios and Palunko, Ivana and De Schutter, Bart},
  journal = {Scientific Data},
  volume  = {11},
  number  = {1},
  pages   = {921},
  year    = {2024},
  doi     = {10.1038/s41597-024-03759-2}
}

@inproceedings{islam2020suim,
  title     = {Semantic Segmentation of Underwater Imagery: Dataset and Benchmark},
  author    = {Islam, Md Jahidul and Edge, Chelsey and Xiao, Yuyang and Luo, Peigen and Mehtaz, Muntaqim and Morse, Christopher and Enan, Sadman Sakib and Sattar, Junaed},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages     = {1769--1776},
  year      = {2020}
}

@article{hong2020trashcan,
  title   = {TrashCan: A Semantically-Segmented Dataset towards Visual Detection of Marine Debris},
  author  = {Hong, Jungseok and Fulton, Michael and Sattar, Junaed},
  journal = {arXiv preprint arXiv:2007.08097},
  year    = {2020}
}

@article{li2025uiis10k,
  title   = {Advancing Marine Research: {UWSAM} Framework and {UIIS10K} Dataset for Precise Underwater Instance Segmentation},
  author  = {Li, Hua and Lian, Shijie and Li, Zhiyuan and Cong, Runmin and Li, Chongyi and Yang, Laurence T. and Zhang, Weidong and Kwong, Sam},
  journal = {arXiv preprint arXiv:2505.15581},
  year    = {2025}
}