Fig. 1: Hierarchical 3D-motion learning framework for animal behavior analysis.

a Data preparation: (1) image streams captured from four cameras with different 2D views; (2) animal body parts are tracked to generates separate 2D skeletal trajectories (color-coded traces); (3) reconstructing 3D body skeleton by integrating these four data streams. b Two-stage NM decomposition to generate the feature space: (1) pose decomposition groups continuous skeleton postural data into discrete postural sequences); (2) NM decomposition, two highlighted (green and orange) blocks represent two NMs decomposed from the postural sequences; (3) NM sequences mapped to their 2D features space (right), where each dot on the 3D axis corresponds to the NM block on the left. c Calculation of locomotion dimension. The continuous velocity of the behaving animal is first calculated, then average the velocity of each segment obtained in the NM decomposition step. d 3D scatter plot represents the combined NM and locomotion feature space. All the movements are clustered into three types (red, green, and orange dots) with the unsupervised approach.