A study on perceptual similarity of human motion

We perform a user study involving different classes of pre-recorded human motions displayed in abstract form either as stick figures or as point lights. Collecting data on more than 1000 user votes of various triples of short motion sequences asking whether a motion A is perceived to be "more similar" to a reference motion O than B or vice versa, we test for associations with numeric distance measures for human motions described in the literature. Our preliminary hypothesis that perceived similarities using stick figure representations are more highly associated to "joint angle based distance measures" than to "point cloud based distance measures" has to be rejected on grounds of the experimental data. We find that there are higher associations for "point cloud based distance measures" than for "joint angle based distance measures" both for the perceived similarities for point light representations as well as for stick figure representations