Yanxi Liu, Thesis Supervisor Rebecca Jane Passonneau, Thesis Honors Advisor
Keywords:
motion capture data Symmetry Machine Learning
Abstract:
Symmetry exists all around us in the natural world and contains valuable information that aids human perception on a variety of subjects. Group theory and computer vision have been extensively used in detecting static visual symmetry, bridging the gap between machine perception and human visual perception. This work expands the notion of symmetry to the motion capture data of dances and proposes symmetry detection methods based upon a texture representation of motion capture data. The symmetry detection methods draw their theoretical foundation from the similarity between dance texture and frieze pattern in group theory, examining three types of symmetry in frieze group: horizontal reflection, vertical reflection, and two-fold rotation. The correspondence between symmetry in dance texture and dance pattern is visualized and presented in the thesis. Lastly, a set of symmetry features are proposed, and the linear regression based upon them demonstrates statistically significant linear relationships (3e-27 p-value) between dance quality rating and symmetry, and outperforms the joint feature based linear model proposed in previous work. The regression analysis shows that more occurrence and longer duration of symmetry often correlate with lower dance quality ratings.
