Quantifying the pairwise fighting behavior of zebrafish in 3D

Two-body fighting behavior occurs throughout the animal kingdom to settle dominance disputes. Qualitatively, fights appear as a sequence of repeated, stereotyped dynamics that lead, ultimately, to a winner. Quantitatively however, little is known about the shorter-time behaviors that compose a fight and the longer-time dynamics of the dominance decision. We study fighting behavior in pairs of male zebrafish imaged at high spatiotemporal resolution in 3D. We leverage advances in convolutional neural networks to track multiple bodypoints of both fish while maintaining organism identity. We use these bodypoints to uncover key features of contest behavior; the fight onset is apparent through static and dynamic displays, followed by escalating aggression, culminating in distinct posture characteristics indicative of submission and dominance. Our approach offers a new system for the quantitative analysis of strongly-coupled social dynamics which can broadly inform theoretical models of mutual assessment.