FishVIS: Ground-truth video instance segmentation dataset of occluded fish on conveyor belt

Source
DataCite
Published
2026
License
Creative Commons Attribution 4.0 International
Themes
Machine Learning & AI, Fisheries & Aquaculture, Agriculture & Farming

Get the data from DataCite

About this dataset

We introduce a novel methodology for collecting ground-truth datasets for video instance segmentation in fisheries. Catches are recorded with EM cameras under various occlusion levels to simulate realistic scenarios onboard a demersal trawler equipped with a conveyor belt. The method relies on RFID tags to precisely track and store information for each individual catch item in the videos. The collected dataset consists of 1,513 specimens across 20 species caught in the North Sea and contains information on species and lengths at the level of each individual, making the dataset ideal as a supplement to existing training datasets or as a stand-alone benchmark dataset for comparing different deep learning models.The dataset is the first in a series of datasets to be published as part of the Horizon Europe project named OptiFish (www.optifish.eu). Please see the readme-file for further...

How the data was collected

Catches are recorded with EM cameras under various occlusion levels to simulate realistic scenarios onboard a demersal trawler equipped with a conveyor belt. The collected dataset consists of 1,513 specimens across 20 species caught in the North Sea and contains information on species and lengths at the level of each individual, making the dataset ideal as a supplement to existing training datasets or as a stand-alone benchmark dataset for comparing different deep learning models.The dataset is the first in a series of datasets to be published as part of the Horizon Europe project named OptiFish (www.optifish.eu).

Browse sustainability datasets on SusanHub