Abstract
Form analysis is fundamental to shape perception and biological motion perception, and is notoriously robust in the face of sparse visual input, such as remote markers located randomly along a shape’s boundary (Kellman & Shipley, 1991), or located at body joints to form point-light actors (Johansson, 1973). A few studies have put orientation and spatial cues into conflict for static shape perception (Day & Loffler, 2009; Levi & Klein, 2000), demonstrating sophisticated interactions between these cues for global form analysis. Here we create animations of sparsely-sampled, dynamic objects comprised of oriented Gabor patches to explore the relative contribution of position and orientation cues under varying conditions of spatial and orientation uncertainty. Frame-by-frame, a small number of points (2, 4 or 6) were randomly sampled along the shape of a human walker (leftward, rightward) or a rotating square (CW, CCW). We used a one-frame lifetime random sampling technique (Beintema & Lappe, 2006) and manipulated Gabor orientation to either coincide with the spatially-defined shape (e.g. leftward walker, CW rotation) or represent the shape with opposite dynamics (e.g. rightward walker, CCW rotation). Spatial frequency was varied (.25, .75, 1.25 cyc/deg) to manipulate the reliability of orientation cues, and Gabor extent (size) was varied to manipulate reliability of spatial cues. We found significant interactions whereby spatial cues dominated perception as orientation uncertainty increased and vice versa. We introduce a "weak fusion" probabilistic model that uses Bayesian inference within local modules to estimate object dynamics (e.g. rotation and walking direction) independently from spatial cues (‘position labels’) or orientation cues. With a single free parameter representing the relative weight of these cues during cue integration, we replicate several important features of the data. These results suggest that form analysis integrates competing position and orientation cues in a simple and statistically near-optimal fashion for perceiving dynamic sampled shapes.
Meeting abstract presented at VSS 2013