J Neurophysiol (November 1, 2002). 10.1152/jn.00289.2002
Submitted on 17 April 2002
Accepted on 28 June 2002

Global Contour Saliency and Local Colinear Interactions

The Rockefeller University, New York, New York 10021

Li, Wu and Charles D. Gilbert. Global Contour Saliency and Local Colinear Interactions. J. Neurophysiol. 88: 2846-2856, 2002. Our visual system can link components of contours and segregate contours from complex backgrounds based on geometric grouping rules. This is an important intermediate step in object recognition. The substrate for contour integration may be based on contextual interactions and intrinsic horizontal connections seen in primary visual cortex (V1). We examined the perceptual rules governing contour saliency to determine whether the spatial extents of contextual interactions and horizontal connections match those mediating saliency. To quantify these rules, we used stimuli composed of randomly oriented nonoverlapping line segments. Salient contours within this complex background were formed by colinear alignment of nearby segments. Contour detectability was measured using a 2-interval-forced-choice design. Contour detectability deteriorated with increasing spacing between contour elements and improved as the number of colinear line elements was increased. At short contour spacing, the detectability reached a plateau with alignment of a few line segments that together formed a contour subtending several visual degrees. At intermediate spacing, saliency built up progressively with a greater number of colinear lines, extending up to 30°. When contour spacing was beyond a critical range (about 2°), however, the detectability dropped to chance levels, regardless of the number of colinear lines. Contour detectability was found to be a function not only of the relative spacing of contour elements with respect to the noise elements but also of the average density of the overall pattern. Furthermore, training significantly improved contour detection, increasing the critical spacing of line elements beyond which contours were no longer detectable. Our data suggest that global contour integration is based on mechanisms of limited spatial extent, comparable to the interactions observed in V1. These interactions can cascade over larger distances provided the spacing of stimulus elements is kept within a limited range.