A unified operator for contour interpolation

Abstract

Heitger et al (1998) proposed a neural model of illusory contour formation based on inputs from contour junctions and real edge orientations. A number of recent results suggest a common mechanism for at least some aspects of illusory and occluded contour formation. These include the existence of hybrid (quasimodal) contours and cases in which interpolation can be shown to precede determination of the modal or amodal appearance of contours (e.g., the Petter effect). Moreover, ecological considerations suggest that coping with occlusion (amodal completion) is likely the primary function of interpolation processes. The Heitger et al model explicitly does not address amodal completion, although some amodal interpolation appears to occur in some model outputs. Here we propose an alternative model designed to handle illusory and occluded contour interpolation in a unified manner. We replace both grouping operators with a single new operator that takes inputs similar to those of Heitger et al (1992) derived from grayscale images and returns a local map of interpolated contours. We compared the models on a wide range of modal and amodal interpolation displays. The revised model performed comparably to that of Heitger et al (1998) on illusory contour displays and better on occluded, quasimodal, and mixed-boundary-assignment displays. It appears that a neural model providing a common basis for illusory and occluded contour interpolation is both plausible and consistent with many phenomena. We discuss several issues within the context of these models (e.g., how best to account for contour interpolation perpendicular to line ends and local boundary assignment) as well as more general issues, such as the role of higher-order operators and the ultimate limitations of models that operate solely on local features, uninformed by more global image or object constraints.

Footnotes

 This research supported by NEI R01 EY013518 (007089-01)