An underwater photograph may contain many animals while its labels name only the ones a specialist was studying. FathomNet’s 2026 detection challenge tackles the resulting trap: a missing label does not establish that an animal is absent.[1]
The challenge supplies 6,463 training images containing both labelled and unlabelled organisms, then evaluates detection against 1,425 fully annotated test images. It consolidates the database’s categories into 32 groups. This is positive-unlabelled learning: known examples are available, but other positive examples remain hidden among unmarked regions.[1]
Imagine a jellyfish labelled in a photograph containing an unmarked fish. Treating everything outside the jellyfish’s box as empty would teach the wrong lesson.[1]
A related task is to make existing annotations more useful. In a July 2025 study, researchers supplied Meta’s Segment Anything Model with previously recorded points and boxes from four marine datasets. The model proposed object boundaries; it did not assign species names.[2]
The results varied with the scene. Crowded organisms could merge into one outline, while shadows could be included with an animal. Relatively clear subjects against a dark midwater background were easier. The study supports assisted annotation with expert checking, and does not establish a measured saving in annotators’ working time.[2]
This separates two jobs that are easily blurred: locating an organism and identifying it. A neat outline can help prepare an image for training while leaving the biological judgement unresolved.[2]
FathomVerse supplies another route to better labels through community participation. Its June 2026 update introduced Filter, a mini-game for sorting and matching images. A Denizen Mode lets experienced participants help verify community annotations, working alongside taxonomy expertise.[3]
At that update, MBARI reported more than 55,000 users and 28 million annotations since the app’s 2024 launch, yielding over 100,000 annotated images shared through FathomNet. Those figures describe different stages of the process: many annotation actions can contribute to one usable image.[3]
The common lesson is practical: preserve what each annotation establishes. A location, an identification and an unchecked region carry different information.[1][2][3]
Sources & further reading
- FathomNetCLEF2026 challenge documentation FathomNet · Accessed 29 Oct 2026
- Assisting human annotation of marine images with foundation models Frontiers in Marine Science · 24 Jul 2025 · Accessed 29 Oct 2026
- MBARI and FathomNet partners release new version of FathomVerse app MBARI · 3 Jun 2026 · Accessed 29 Oct 2026