New iterative consensus protocol allows for more teamwork and faster localization

Dear FathomNauts,

We’re making changes to Bound to generate localizations faster, enabling FathomNauts to process image data quicker. To explain this update, let’s break down how Bound used to work. The goal of Bound is to box the objects of interest, including animals, equipment, and trash, in each image. This is called localization and it is the first step in annotating data for AI model training.

In the Bound mini-game, pink boxes represent AI predictions.

When you enter Bound, you are presented with an image that usually contains some pink boxes. Those pink boxes are generated by our AI model, which predicts where objects of interest are located in the image. Then, it’s up to you to edit the boxes provided and add more boxes around animals or objects in the frame. On the backend of FathomVerse, we run a community consensus protocol that gathers all the bounding boxes FathomNauts have submitted and determines which boxes the community agrees on. In other words, every box in Bound acts as a vote towards consensus, and when there are enough similar votes, we can be confident that an image is properly localized.

Since most data have to be fully localized in Bound before moving on to classification, the faster we move imagery through this stage the better. That's why we're changing how the consensus protocol works.

We are implementing a change to accelerate localization consensus by using an iterative consensus protocol. Whereas before the pink boxes in Bound represented AI predictions, now the pink boxes are also going to include player-generated boxes; together these are considered proposals. This means that when you play Bound you could see a bounding box that another FathomNaut created and edit your fellow FathomNaut’s localization. This new iterative consensus protocol will allow FathomNauts to work together to refine boxes, completing the localization stage faster, and accelerate sharing of annotated images with researchers.

What do you need to do differently? Nothing. We just want to be transparent about the data we are showing you. You might see more box proposals in Bound than before. Please keep refining existing boxes, removing bad ones, and adding new ones to tightly and accurately box all animals, equipment, and trash in each image. We are gradually implementing this new approach so it won’t apply to all the images right away. As we monitor the new consensus protocol for Bound, please let us know if you notice any problems. The best way to notify our team of bugs is on the FathomVerse Discord.

Onwards and downwards,

Lilli

FathomVerse Engagement Coordinator

fathomverse@mbari.org