Abstract
Movements of the human body are comprised of multiple motion components organized in a hierarchical manner. For example, the motion of a foot joint is nested within a moving leg, which is nested within a moving body. It remains unknown how humans discover the hierarchical structure representing bodily movements, and what fundamental constraints underlie the formation of motion structure. A non-parametric Bayesian model was developed to infer the hierarchical motion structures that can generate observed visual input. Many different hierarchical motion structures would be consistent with observed bodily movements, but three types of constraints sufficed to resolve ambiguity: a generic prior on object parsing (favoring simpler hierarchical structures), a generic prior on spatially smooth motion within a component (resulting from rigid movements of individual parts), and a body-specific constraint (certain joints move in coordination based on body structure). This generative model can account for a range of findings in the literature on biological motion perception, including better recognition performance for an upright than an inverted point-light walker, above-chance performance in discriminating walking direction for a spatially-scrambled and phase-scrambled walker (Troje & Westhoff, 2006, Exp. 2), the special contribution of foot movements in discriminating walking direction (Troje & Westhoff, 2006, Exp. 3), the variation in recognition accuracy across different action categories (Dittrich, 1993). The model also reveals how body structure interacts with the organization of motion components to form efficient action representations. When the body constraint was included in the formation of hierarchical structure, the representation space of actions was altered to cluster similar action types, but enlarge the distances of different actions. This finding suggests that the body constraint plays an important role in organizing motion information into a hierarchical structure, thereby increasing the representational power of an action space.
Meeting abstract presented at VSS 2016