Multimodal Information Fusion for Automatic Aesthetics Evaluation of Robotic Dance Poses

Abstract

Aesthetic ability is an advanced cognitive function of human beings. Human dancers in front of mirrors estimate the aesthetics of their own dance poses by fusing multimodal information (visual and non-visual) to improve their dancing performances. Similarly, if a robot could perceive the aesthetics of its own dance poses, the robot could demonstrate more autonomous and humanoid behavior during robotic dance creation. Therefore, we propose a novel automatic approach to estimate the aesthetics of robotic dance poses by fusing multimodal information. From the visual channel, the shape features (including eccentricity, density, rectangularity, aspect ratio, Hu-moment Invariants, and complex coordinate based Fourier descriptors) are extracted from an image; from the non-visual channel, joint motion features are obtained from the internal kinestate of a robot. The above two categories of features are fused to portray completely a robotic dance pose. To automatically estimate the aesthetics of robotic dance poses, the following ten machine learning methods are deployed: Naive Bayes, Bayesian logistic regression, SVM, RBF network, ADTree, random forest, voted perceptron, KStar, DTNB, and bagging. Experimental results show the feasibility and good performance of the proposed mechanism, which was implemented in a simulated robot environment. The highest correct ratio of aesthetic evaluation is 81.6%, which comes from the ADTree, based on the above mixed features (joint + shape).