Evaluating accounts of human contour integration using psychophysical and computational methods

Abstract

Field, Hayes and Hess (1993) offered path-paradigm (PP) stimuli that enabled the investigation of the Gestalt processes of proximity and good continuation without the potential confounds present in earlier studies. They proposed that only an association field, a mechanism able to integrate dynamically the outputs of filters with different orientation preferences, could detect their stimuli. This thesis describes simulations which examined whether the PP task could be solved without recourse to an association field. A simple-filter model (SFM) was tested, were each response image was the result of the convolution of the stimulus image with an oriented filter. The lengths of zero-bounded regions (ZBRs) within each image were calculated (Watt, 1991). In a simulated 2AFC trial the status of target was assigned to the image containing the longest ZBR. Results confirmed the Hess and Dakin (1997) finding that Field et al’s PP stimuli could be successfully detected by the SFM. Further simulations conflicted with Hess and Dakin; the SFM was also able to detect stimuli containing phase-alternated contours. Thus, it is not necessary to invoke an association field mechanism to explain contour integration, even for phase-alternated stimuli.

Psychophysical experiments indicated that the closure effect reported by Kovaks and Julesz (1993) may actually be caused by contour smoothness, rather than closure per se. Where local properties are held constant, manipulations of contour closure are inevitably confounded by smoothness. Further modelling experiments revealed that, for the SFM, detection rates were inversely correlated with global smoothness, a reversal of the pattern found for human observers. This phenomenon provides a useful means of investigating the PP task in the peripheral field. Hess and Dakin (1997) have argued that the SFM is sufficient to account for contour integration processes in the periphery. Thus, for the periphery there should be an inverse relationship between smoothness and detectability. Experimental results revealed that this is not the case, smooth contours were detected more frequently than jagged contours. It is concluded that the Field et al. (1993) conception of the association field was largely correct, despite the fact that their stimuli were potentially flawed. However, the association field model may need modification in order to account to the effect of contour smoothness. It is suggested that the association field model should be considered a component in Roelfsema, Lamme and Spekreijse’s (2000) model of curve-tracing. Within this framework the process of ‘chunking’ (Mahoney and Ullman, 1988) might explain the improved detection of frequently experienced contour shapes, i.e. straight lines and smooth arcs. Furthermore, it is suggested that this mechanism may be acquired through visual experience and that it could underlie PP performance in the whole visual field.