Abstract
Underestimation of slant is a well-known bias in human surface attitude perception. Does this frontoparallel bias have a rational basis? If we consider a single tilt axis, and assume a world in which surfaces are uniformly distributed across slant, then a frontoparallel bias is predicted: surfaces with smaller slants project to larger regions of the visual field than more slanted, foreshortened surfaces and therefore should be observed more frequently. However, in natural scenes, surface slant is non-uniformly distributed and surface tilt varies widely. Here we consider the joint estimation of slant and tilt and relate it to natural scene statistics. On each trial, observers viewed a real, textured surface presented at one of 36 different orientations. The surface was viewed monocularly, through a small aperture, to reduce the reliability of slant and tilt cues, and therefore maximise the effect of biases related to perceptual priors. The use of a real surface eliminated cue conflicts that occur when slanted stimuli are rendered on a frontoparallel screen, which could otherwise lead to underestimation of slant. Observers reported perceived slant and tilt by rotating an unseen haptic paddle to match the orientation of the viewed surface. Consistent with previous reports of a frontoparallel prior, our results suggest a strong tendency to underestimate surface slant. However, slant underestimation was greater for ceiling planes than wall or floor planes. Evidence for a bias toward floor planes is also apparent in the tilt data, and reversals of perceived tilt were more frequent for ceiling planes than ground planes. Analyses of natural scenes statistics from the Southampton-York Natural Scenes (SYNS) dataset suggest that these perceptual biases are well-matched to the distribution of surface slant and tilt in natural scenes.
Meeting abstract presented at VSS 2015