Abstract
In perceptual experiments, the perceived slant is frequently less than the physically specified slant. We show that this slant underestimation can be successfully predicted by a probabilistic model that combines current measurements with a prior expectation of zero slant. From the geometry of natural scenes, one can make a prediction about the shape and spread of the prior. The probability that a line of sight will intersect a surface rotated by a random amount about a vertical axis is a half-cosine distribution centered at 0deg. We asked whether the visual system behaves as if it has internalized this predicted distribution of likely slants. The probabilistic model predicts different effects of the prior depending on the reliability and slant of the surface. Estimating the visual system's internal prior from measured psychophysical data is challenging because observer responses are affected by both the prior and the unknown likelihood. To measure the visual system's prior, we presented two slanted planes and observers indicated which was more slanted. The stimuli were either regular grid-like textures for which slant was reliably discriminated (R) or irregular textures for which slant was unreliably discriminated (U). A range of base slants was presented in three types of trials: UvsU, RvsR, and UvsR. To eliminate any cues that could bias the visual system towards slant estimates of zero, we used real surfaces. We found that observers systematically underestimated the slant of the unreliable stimulus relative to the reliable stimulus in a manner consistent with the probabilistic model. From the psychophysical data we inferred the spread and shape of the internal prior distribution using a technique similar to Stocker and Simoncelli (2006). The priors reconstructed from the data were peaked at zero slant and were similar to the theoretical expectation.
NIH, NSF, AOF Ezell Fellowship for JB.