Abstract
The perception of boundary shape depends upon the organization of local orientation signals into global contours. Models generally assume that grouping is based upon local Gestalt relationships such as proximity and good continuation. While there have been reports that the global property of contour closure is involved in this process (e.g., Kovacs & Julesz 1993), a more recent study suggests otherwise (Tversky, Geisler & Perry 2004). This raises the question: is contour grouping completely insensitive to global properties of the stimulus, depending only upon local Gestalt cues?
To address this question, we conducted a psychophysical experiment in which observers were asked to detect briefly-presented target contours in noise. Contours were represented as sequences of short line segments, and the noise was composed of randomly positioned and oriented segments of the same length. We used QUEST to estimate the threshold number of noise elements at 75% correct performance in a present/absent task.
Three conditions were tested. In Condition 1, targets were the closed bounding contours of 391 animal shapes derived from the Hemera object database. These contours afford local Gestalt properties but also a host of global properties, including closure. In Condition 2, we created first-order metamers of these contours by randomly shuffling the order of the angles between neighbouring segments. This preserves all local Gestalt properties exactly, but destroys all higher-order properties. In Condition 3, we also randomized the signs of the angles, thus removing a convexity bias.
While noise thresholds were similar for Conditions 2 and 3, they were significantly higher for Condition 1, suggesting a global influence on grouping. Further analysis suggests that this difference cannot be explained by differences in stimulus eccentricity, element density, or contour intersections produced in shuffled stimuli. Instead the results point to a process of perceptual organization that goes beyond local, first-order cues.
This work was supporte by grants from NSERC and GEOIDE