UC Davis

This work aims to apply advancements in deep learning for image classication to improvethe recognition of movement style in motion capture data. A RESNET architecture is used to classify individual dancers based on clips of their movement and to predict style based on clips of various motions in 7 dierent style categories -angry, childlike, depressed, proud, etc. Motion capture clips from trained dancers at George Mason University performing the same choreographic sequence several dierent times were used for a dancer identication task. A data set of actions performed with dierent labeled styles such as proud, depressed, angry, old, and childlike created by [39] was used for a style identication task. Results were compared using Quaternion, scaled positional coordinates, and Euler angle representations of the motion capture clips supplied to the network for learning.