Abstract
The visual world is full of contours, with a preponderance of cardinal (vertical and horizontal) orientations (Switkes et al., 1978). Do observers behave in a Bayesian fashion, with visual orientation estimates based on a prior distribution reflecting these statistics? For such a Bayesian estimator, estimated orientation of ambiguous stimuli will be biased toward cardinal orientations. Methods: On each trial, two stimuli were shown, one on the left and one on the right side of the display, for 750 ms. Each consisted of 64 Gabor patches in a circular window. The Gabor orientations in a stimulus were either all identical (L: low noise) or were normally distributed with standard deviation approximately 20 deg (H: high noise; SD chosen per observer based on a pilot discrimination experiment). One stimulus (the standard) had a fixed mean orientation (varied across blocks); the other (the comparison) varied in mean orientation. Observers indicated which stimulus had greater (more “clockwise”) mean orientation. Conditions (across blocks) included 12 standard orientations (spanning 180 deg) and three comparison types (LvL, HvH, LvH). The LvL and HvH conditions allow one to estimate the just-noticeable difference as a function of orientation at each noise level. The LvH condition allows one to estimate the prior distribution used by observers (Stocker & Simoncelli, 2006). Results: Observers showed a classic oblique effect—better discrimination at the cardinals—in the LvL conditions, but not in the HvH conditions due to stimulus noise. In the LvH conditions, observers behaved as if the perceived orientation of the high-noise stimulus was systematically biased toward the nearest cardinal orientation. The data are consistent with a Bayesian observer that uses a non-uniform prior distribution over orientation, with peaks at the cardinal orientations.