FathomNet: A global image database for enabling artificial intelligence in the ocean

Area of Science:

• Marine biology
• Oceanography
• Computer science

Background:

• The ocean faces rapid environmental changes, necessitating effective visual monitoring of marine life.
• Current data collection and analysis capabilities struggle to keep pace with the required spatiotemporal scales for ocean stewardship.
• Machine learning for visual data analysis has been limited in marine environments due to data standardization and labeling challenges.

Purpose of the Study:

• To develop an open-source image database, FathomNet, for standardizing and aggregating curated, labeled marine visual data.
• To facilitate the training and deployment of machine learning models for analyzing underwater imagery.
• To enable automated tracking of underwater objects and concepts, supporting ocean research and conservation.

Main Methods:

• Creation of FathomNet, an open-source database, to standardize and aggregate labeled marine imagery.
• Seeding FathomNet with diverse imagery including marine animals, equipment, and debris.
• Demonstration of using FathomNet data to train machine learning models and reduce annotation efforts.
• Integration of FathomNet with robotic vehicles for automated tracking.

Main Results:

• FathomNet provides a standardized, curated dataset for marine visual data analysis.
• Trained models using FathomNet data can be deployed on other video datasets, reducing manual annotation needs.
• Automated tracking of underwater concepts is achievable when FathomNet is integrated with robotic systems.

Conclusions:

• FathomNet accelerates the processing of visual data for ocean research and stewardship.
• The database supports the development of machine learning applications for marine environments.
• Community contributions to FathomNet will enhance its utility for achieving a healthy and sustainable global ocean.