Abstract
Human observers readily recognize complex actions based on impoverished visual inputs, such as point-light biological motion. How is point-light biological motion represented by the visual system? We used techniques of classification images to provide a pictorial answer. Observers discriminated a point-light human who was walking either forward or backward (moonwalk). The two stimulus movies were identical except that one of them was played backward. The point-light human was embedded in dynamic white noise. We developed a new method for constructing classification images by computing trial-by-trial correlations between noise pixels and an observer's responses. The resulting correlation map revealed statistically significant correlations at all point-light locations, i.e., dynamic “templates” of the forward and backward walkers. We further computed a semipartial correlation map, using multiple regression to increase the power of the analysis in order to reliably detect small but nonzero correlations between noise pixels and responses. Classification movies were created using the resultant semipartial correlation map, in which noise pixels located along the “virtual skeleton” (lines connecting point-light joints according to the structure of a human body) yielded significant correlations with observers' responses. In a control analysis, lines linking up symmetrically corresponding point-light joints (e.g., the left and right hands) yielded non-significant correlations.
Our findings indicate that the internal representation used to discriminate dynamic point-light walkers includes not only joint positions (conveyed by the physically-presented point-lights), but also the “skeleton” that was never physically present in the stimulus signals.