Abstract

Our understanding of collective animal behavior is limited by our ability totrack each of the individuals. We describe an algorithm and software,idtracker.ai, that extracts from video all trajectories with correct identitiesat a high accuracy for collectives of up to 100 individuals. It uses two deepnetworks, one detecting when animals touch or cross and another one for animalidentification, trained adaptively to conditions and difficulty of the video.

 

Quick Read (beta)