FathomNet2024

Start Date: April 23 End Date:  June 13 Competition URL: Kaggle

Motivation

The ocean is one of the least-explored habitats on Earth. It is home to diverse animals, many of which are either unknown to science or poorly understood. Scientists are increasingly leveraging ocean-going camera systems to monitor populations and discover new animals. This combination of dense, image-based sampling and a poorly defined space of classes makes for a challenging computer vision problem: How do you get a machine-learning model to flag new and different animals?

For this competition, we have curated data from the broader FathomNet image set that is emblematic of this challenging use case. The training set contains 18 supercategories of marine animals collected along the ocean floor. The test set contains these same 18 supercategories AND new ones not represented in the training data. The challenge is to develop a model that can recognize the original 18 supercategories and flag unknown objects. Developing novel or general category discoveries for ocean research will help scientists rapidly find unlabeled objects in historical datasets and enable creative field campaigns to seek out new life.

The images for both the training and target data were collected by a single camera system deployed by the Monterey Bay Aquarium Research Institute (MBARI) on several of its Remotely Operated Vehicles off the coast of Central California. Participants should not use other image sets to train their models; however, they can leverage other types of outside data sources (e.g., text descriptions, etc.) as they see fit. Pretrained models can be used for the initialization of training (e.g., ImageNet or COCO classification or detection models provided by many deep learning packages). Teams should plan on specifying additional data sources and/or pre-trained models used when uploading results.

Organizers