Curve-Based Shape Representation in Visual Perception

Shape is the predominant cue for object recognition in visual perception. Though many studies have demonstrated the psychological importance of shape information, much remains unknown about how the visual system forms representations of shape. Shape representations are unlikely to be a literal recording of an object’s boundary. Rather, representations of shape are abstract in that they encode relations between parts, are economical, selectively encoding information present in the physical stimulus, and are invariant to 2D transformations and changes to the properties of local elements. In this dissertation, I examine evidence for the theory that representations of shapes are formed by partitioning a contour into regions of similar curvature and representing segments with a single curvature value. I first develop a computational model for how contours could be recoded abstractly as sets of constant curvature segments. I experimentally tested two free parameters in the model and then tested the model’s ability to predict the perceptual difference between pairs of shapes. In Chapter 2, I showed how the visual system could encode constant curvature representations of shape from activations of oriented luminance contrast detectors in early vision, bridging a theoretical gap between subsymbolic activations that are responsive to light energy and symbolic representations that are concerned with objects, contours, and surfaces. In Chapters 3 and 4, I applied the constant curvature theory to two interesting domains of shape perception. First, I tested how and why people encode shape representations from arrays of unconnected dots. Consistent with the constant curvature theory of shape, dot arrays that were perceived to have curvilinear contours were more easily represented as shapes than dot arrays perceived to have straight edges joined at corners. In Chapter 4, I studied shapes with both global form and high frequency local contour features. Evidence was found for a hypothesis that local and global contour features are encoded independently and in separate systems. In this theory, global features are extracted from large curvature detectors and described in detail while local contour features are extracted from small curvature detectors and encoded with a few descriptive statistics rather than as individual features