Detection of natural shapes in noise

Abstract

We studied the detection of natural shapes embedded in random backgrounds. Detecting coherent objects in cluttered background is a basic function of perceptual organization. Numerous studies have investigated the case of simple open contours, but few have considered the more naturalistic case of complex closed shapes. We studied the detection of natural shapes (the bounding contours of animals and leaves drawn from several natural shape databases) embedded in random backgrounds. In previous work on open contours (VSS2011) we showed that detection performance falls off rapidly with contour complexity, quantified using an information-theoretic measure (cumulative surprisal, Feldman & Singh 2005). For closed contours (enclosing shaped regions), we generalize this complexity measure using the probabilistic shape-generating skeletal model of Feldman & Singh (2006), which provides a simple measure of global (region-based) shape complexity based on description length (DL) conditioned on the estimated skeletal model. In the tested displays, contours consisted of chains of pixels embedded in random monochromatic pixel noise. We evaluated proportion correct (2IFC) as a function of simple contour complexity (integrated surprisal along the contour), the new closed shape complexity measure (cumulative surprisal given the estimated skeletal model), and several conventional measures of global shape properties such as convexity and compactness. Performance degraded markedly with both contour complexity and the new measure of global shape complexity, falling to near chance levels as complexity approached purely random structure (maximum Kolmogorov complexity). The results suggest that both contour and region geometry play important roles in representing form. Meeting abstract presented at VSS 2012