Measuring the Perceptual Similarity between Sets of Two People's Interactive Motions in a Coarse-to-Fine Manner

Description

Evaluating the similarity of motions is useful for motion retrieval, motion blending, and performance analysis of dancers and athletes. However, existing similarity measures are only designed for evaluating single person's motions. The researchers propose a perceptual similarity model that measures sets of two people's interactive motions in a coarse-to-fine manner. In the coarse matching stage, the non-matched interactive motions will quickly be filtered out. In the fine matching stage, the remaining interactive motions will be matched with details using the local information about the interactions. Human perception about the motion similarity will be correlated with the activity level of the interactions among the body parts between the two people during their interactive motion. The weights for the features in this perceptual similarity model will be optimized. An advantage of this proposed model is that there is no need to recognize each individual's action. In addition to the theory part, two applications will be implemented to show off the applicability of this proposed similarity model. The first application is an interactive motion game whose objective is to let two players imitate certain interactive motions and such kind of system has high potential to become the next generation entertainment platform. The second application is a retreival interface that allows a single user to easily define a two people's interactive motion as query and search for similar motions in the database. A database containing different classes of two people's interactive motions will be captured and made available to the public to facilitate future research for animation, retrieval, biomechanics and entertainment.