Abstract
A well known result in psychophysical studies of good continuation in arrays of Gabor patches is that the visual system is better at detecting smooth paths rather than jagged ones, and all the more so as they are formed by elements that are roughly aligned to the local tangent of the contour (association field, Field et al. 1993). Here we present a similar experiment on contour detection, and a stochastic model that predicts and interprets the perceptual thresholds for this task, relying on the non-accidentalness principle. Our experiment consists in an attentive contour detection task in arrays of Gabor patches. The visibility of contours among the cluttered background is affected by three varying parameters: the number of target elements, the amount of angular noise deviating their orientations from the local tangents to the contour, and the total number of patches in the image. Sixteen subjects took the experiment and their detection performance was compared to an artificial observer algorithm, on every stimulus. We built this algorithm on the a-contrario theory (Desoneux et al. 2008), applied here to formalize mathematically the non-accidentalness principle for good continuation and proximity. To predict the salience of curves, it associates with each candidate percept a measure, the Number of False Alarms (NFA), quantifying its degree of masking. The NFA showed a strong correlation with detection performance: different targets with the same NFA yielded similar levels of dectability among subjects. Furthermore, the algorithm's answers matched accurately those of human subjects, on average as well as on a trial-by-trial basis. The overall results give credit to the non-accidentalness principle, as a way to interpret and predict the perceptual grouping in masking conditions. Future work will concentrate on predicting the salience of symmetry using the same framework.
Meeting abstract presented at VSS 2016