FathomNetCLEF2026 @ LifeCLEF & CVPR-FGVC

Positive-Unlabeled Object Detection in Marine Images

Underwater imaging opens a unique window into the deep sea. It is one of the few ways scientists can directly observe marine animals and habitats, creating visual records that are essential for understanding our changing ocean. But humans generate more ocean data than we can handle, and computer vision is vital to process the thousands of hours of video collected by researchers around the world. Compared to terrestrial datasets, marine imagery presents many unique challenges for computer vision, including high taxonomic diversity, visually similar morphologies, and environments that range from benthic habitats to the open midwater.

One of the biggest challenges? Incomplete labels. Marine biologists are experts in specific animals so they only label what they know in each image. Consequently, annotators may carefully focus on jellyfish in a given image, while ignoring the dozen other creatures swimming by. Just because something isn't labeled doesn't mean it isn't there, and that's a serious problem for anyone trying to teach a computer to recognize all ocean life.

Read more