Abstract
Humans are known to rely profoundly on bounding contours for object detection, segmentation and recognition. This task is complicated by the interposition of other objects in the visual field that lead to occlusion: partial blocking of one object by another. However, the impact of occlusion on human perception is mitigated by the human ability to perceptually complete partially occluded bounding contours, i.e., to fill-in the missing shape information. Here we examine the degree to which the human brain uses local and non-local cues to solve this perceptual completion task. Each visual stimulus consisted of a sequence of dots regularly sampling the outline of a 2D shape. To simulate occlusion, a contiguous interval of 10-50% of the dot pattern was extinguished. Observers were asked to adjust a probe dot along a linear axis orthogonal to the gap until the dot appeared to lie where the contour would be, were it visible. Two classes of shape were employed: animal shapes, which afford both local and global cues to completion, and metamer shapes, which match the curvature statistics of the animal shapes but are otherwise random, thus affording local but not global cues to completion. Mean absolute error was lower for animal shapes than for metamer shapes and, while completions tended to be negatively (inward) biased for both, the bias was less for animals than metamers. Local linear and elastica models were better able to account for human completions of metamer shapes than animal shapes. Together these findings point to a contribution of non-local shape cues to perceptual contour completion.