Scalable Self-Supervised Pipelines for Natural Behaviour Modeling in animals
A scalable self-supervised framework that learns 4D representations of natural behaviour
The project develops a scalable self-supervised pipeline to learn 3D behavioural representations from continuous home-cage video, using a combination of 4D motion modelling, unsupervised pose estimation, and sequence prediction. By training across multiple rodent models, we aim to uncover the intrinsic structure of natural behaviour without manual annotations, with the broader goal of building a generalizable foundation model for automated behavioural neuroscience and long-term monitoring.