This web application uses cookies and other tracking technologies to ensure you get the best experience.
SC/69B/PH/07
Download this file5.5 MB |
Download |
View directly in browser5.5 MB |
View in browser |
Resource ID
22195
Access
Open
Document Number
SC/69B/PH/07
Full Title
Flukebook Continuing Advancements in Multispecies AI
Author
Lasha Otarashvili, Jason Holmberg, Tamil Subramanian, Jacob Levenson2
Authors Summary
This paper presents the AI experiments of the authors on re-identifying individuals in the combined photo ID catalogs of 23 species (22 cetacean; 1 shark). The results demonstrate that AI training across species and differing ID presentation (flukes, fins, laterals, etc.) can significantly improve per-species ID prediction accuracy. Incomparison to similar work, this paper presents re-ID of individuals in a novel "open world" approach to AI and leverages a larger dataset and different species composition. It also presents the emergence of "zero shot" learning capability to re-ID individuals in unseen classes/species as well through the formulation of multispecies, single species, and leave-one-out experiments. The resulting AI model is now deployed for 34 species for Flukebook.org users and has inspired similar results by the same authors for wildlife face ID and terrestrial carnivore re-ID.
Publisher
IWC
Publication Year
2024