The quantification of animal social behaviour is an essential step to reveal brain functions and psychiatric disorders during interaction phases. While deep learning-based approaches have enabled precise pose estimation, identification and behavioural classification of multi-animals, their application is challenged by the lack of well-annotated datasets. Here we show a computational framework, the Social Behavior Atlas (SBeA) used to overcome the problem caused by the limited datasets. SBeA uses a much smaller number of labelled frames for multi-animal three-dimensional pose estimation, achieves label-free identification recognition and successfully applies unsupervised dynamic learning to social behaviour classification. SBeA is validated to uncover previously overlooked social behaviour phenotypes of autism spectrum disorder knockout mice. Our results also demonstrate that the SBeA can achieve high performance across various species using existing customized datasets. These findings highlight the potential of SBeA for quantifying subtle social behaviours in the fields of neuroscience and ecology.

Multi-animal behaviour quantification is pivotal for deciphering animal social behaviours and has broad applications in neuroscience and ecology. Han and colleagues develop a few-shot learning framework for multi-animal 3D pose estimation, identity recognition and social behaviour classification.