Semi-supervised animal action segmentation: head-fixed mouse

This is a publicly available dataset from the International Brain Lab (IBL et al. 2023, bioRxiv). A head-fixed mouse performed a perceptual decision-making task, using a wheel to indicate its decision (IBL et al. 2021, eLife; IBL et al. 2023, bioRxiv). Two cameras – “left” (60 Hz) and “right” (150 Hz) – capture roughly orthogonal side views of the mouse’s face and upper trunk during each session (IBL et al. 2022, Figshare), though we only use the left view.Multiple points are tracked across the paws and face, though for this work we only consider a single tracked point on the paw closest to the camera. In addition to the paw coordinates, we also use the 1D wheel velocity as input to the models.We consider four behavioral categories for this single paw: still, moving without turning the wheel, moving while turning the wheel, and grooming. A single annotator labeled behaviors using the DeepEthogram GUI (Bohnslav et al. 2021, eLife). We use five labeled videos for training, and five for testing. The dataset contains 1M frames, 14k of which are labeled.