vapaau/autofish

The AUTOFISH dataset comprises 1500 high-quality images of fish on a conveyor belt. It features 454 unique fish with class labels, IDs, manual length measurements, and a total of 18,160 instance segmentation masks.

The fish are partitioned into 25 groups, with 14 to 24 fish in each group. Each fish only appears in one group, making it easy to create training splits. The number of fish and distribution of species in each group were pseudo-randomly selected to mimic real-world scenarios.

Every group is partitioned into three subsets: Set1, Set2, and All. Set1 and Set2 contain half of the fish each, and none of the fish overlap or touch each other. All contains all the fish from the group, purposely placed in positions with high overlap. Every group contains 20 images for each set, where variation is introduced by changing the position and orientation of the fish. Half the images of a set are with the fish on one side, while the other half has the fish flipped. This structure can be seen in the dataset viewer*.

The following figures display some examples with overlaid annotations:

The available classes are:

• Cod
• Haddock
• Whiting
• Hake
• Horse mackerel
• Other

Other information contained in the annotations:

• Segmentation masks
• Bounding boxes
• Lengths
• Unique fish IDs
• 'Side up' referring to the side of the fish that is visible

In addition to all the labeled data, two high-overlap unlabeled groups, as well as camera calibration images are included.

You can load this dataset with a default split configuration using the datasets library

dataset = datasets.load_dataset('vapaau/autofish', revision='script', trust_remote_code=True)

If you use this dataset for your work, please cite:

@misc{bengtson2025autofishdatasetbenchmarkfinegrained,
      title={AutoFish: Dataset and Benchmark for Fine-grained Analysis of Fish}, 
      author={Stefan Hein Bengtson and Daniel Lehotský and Vasiliki Ismiroglou and Niels Madsen and Thomas B. Moeslund and Malte Pedersen},
      year={2025},
      eprint={2501.03767},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2501.03767}, 
}

Ethical Statement

Fish used in these experiments were caught and landed by fishermen following relevant legislation and normal fishing procedures. The Danish Ministry of Food, Agriculture and Fisheries of Denmark was contacted before fish collection to ensure compliance with legislation. The fish were dead at landing and only dead fish were included in this experiment. There is no conflict with the European Union (EU) directive on animal experimentation (article 3, 20.10.2010, Official Journal of the European Union L276/39) and Danish law (BEK nr 12, 07/01/2016). The laboratory facilities used at Aalborg University are approved according to relevant legislation.

---

*Due to size limitations we chose to display 2 random groups on the dataset viewer instead of the entire dataset.