DUMAL

An Underwater Megafauna Detection Dataset for Sustainable Longline Fisheries

Esteban Darizcuren-Dreau¹ Nicolas Sidere¹ Jean-Christophe Burie¹ François Poisson²

¹L3i, La Rochelle Université, France    ²MARBEC, Univ Montpellier, CNRS, IFREMER, IRD, Montpellier, France

WACV 2027

[Dataset] [Code] [BibTeX]

Abstract

Longline fishing poses significant bycatch risks to protected marine species, yet no public computer-vision dataset records megafauna in situ around active fishing gear. We introduce DUMAL (Detecting Underwater Megafauna Around Longline fishing gear), a dataset of 60,870 annotated frames (94,101 boxes across nine species) from 55 independent videos recorded on an instrumented pelagic longline in the Mediterranean Sea. Two inter-annotator agreement studies validate annotation quality: a localization study (mean IoU 0.605 detection-and-localization, 0.811 localization-only, within four points of the human ceiling) and a classification study (Cohen's κ = 0.86). We split train and test hermetically at the video level, a practice uncommon in underwater detection datasets. Benchmarking eight detector configurations, we find that generalization to unseen footage depends far more on the number of distinct training videos a class appears in than on its instance count, confirmed by two controlled ablations that independently vary video count and frame density.

This dataset is part of the MEGABIT project.

Data acquisition

Cameras were mounted on a subset of the branch lines of a pelagic longline (4 to 38 cameras per fishing operation), facing downward 1–2 m above the baited hook. Footage was recorded between 2019 and 2025 off Corsica and in the Gulf of Lions (France), by day and by night, during regular commercial fishing operations. From about 500 hours of raw footage, a marine biologist selected the sequences showing animal activity, which gave the 55 annotated source videos. Three camera models were used, so images are 1280×720, 1920×1080 or 2560×1440.

Instrumented pelagic longline. On some branch lines, a camera is mounted 1–2 m above the hook.

Statistics

Train and test are split at the video level. Videos recorded by the same camera on the same day are grouped and always kept in the same split. Validation frames (about 10%) are drawn from the training videos.

Jellyfish appears in a single video, so it has no test instances and is not evaluated.

Benchmark

Results of the eight detector configurations on the test videos. See the paper for per-class results and the ablation studies.

Download

The dataset (13.7 GB) can be downloaded directly from the links below. The same files are archived on Harvard Dataverse: https://doi.org/10.7910/DVN/QM19AI. It is released under the CC BY 4.0 license.

• DUMAL_annotations.zip: COCO and YOLO annotations, data.yaml, ablation splits (28 MB)
• Train images (11.1 GB): part001 (2.5 GB), part002 (2.5 GB), part003 (2.5 GB), part004 (2.5 GB), part005 (1.2 GB)
• val_images_part001.zip: val images (1.2 GB)
• test_images_part001.zip: test images (1.3 GB)

Put all the files in the same directory and run:

unzip DUMAL_annotations.zip
unzip 'train_images_part*.zip' -d DUMAL/train/images
unzip val_images_part001.zip -d DUMAL/val/images
unzip test_images_part001.zip -d DUMAL/test/images

This gives:

DUMAL/
├── data.yaml
├── annotations/        instances_train.json, instances_val.json, instances_test.json
├── train/images/  train/labels/
├── val/images/    val/labels/
├── test/images/   test/labels/
└── ablations/
    ├── ablation_video/
    └── ablation_frame/

COCO annotations use category ids 1–9, YOLO labels use class ids 0–8 (see data.yaml). Frames named *_bg_* are background frames without any animal (2,185 frames, no annotation). The ablations/ folder contains the reduced training sets used in the two ablation studies of the paper.

Citation

Any use of this dataset requires citing the following paper:

@inproceedings{darizcuren2027dumal,
  title     = {{DUMAL}: An Underwater Megafauna Detection Dataset for Sustainable Longline Fisheries},
  author    = {Darizcuren-Dreau, Esteban and Sidere, Nicolas and Burie, Jean-Christophe and Poisson, Fran{\c{c}}ois},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027}
}

Acknowledgments

The at-sea experiments were conducted as part of two research programs: the Gen&Rec project ("Distribution and behaviour of swordfish spawners and juveniles in the vicinity of Corsica"), funded by the European Union (European Maritime and Fisheries Fund, EMFF) and the fishing industry (France Filière Pêche, FFP); and the MOUVBLUE project ("Movements of blue sharks"), funded by the Office Français de la Biodiversité (OFB) and the France Relance plan.

Contact

Esteban Darizcuren-Dreau: esteban.dreau-darizcuren@univ-lr.fr
In case of any problem, please contact "l3i-pn$@$univ$-$lr$.$fr" (without $).