Starry night: a texture devoid of depth cues.

From a modern Bayesian point of view, the classic Julesz random-dot stereogram is a cue-conflict stimulus: Texture cues specify an unbroken, unslanted surface, in conflict with any variation in depth specified by binocular disparity. We introduce a new visual texture-the starry night texture (SNT)--that is incapable of conveying slant, depth edges, or texture boundaries, in a single view. For SNT, changing density is equivalent to changing intensity, so an instance of the texture is characterized (up to the random locations of the texture elements) by what we call its densintensity. We consider deviations from the ideal that are needed to realize the texture in practice. In three experiments with computer-generated stimuli we examined human perception of SNT to show that (1) the deviations from the ideal that were needed to realize SNT do not affect the invariance of its appearance across changes in distance of several orders of magnitude; (2) as predicted, observers match SNT across changes in distance better than other textures; and (3) the use of SNT in a slant perception experiment did not reliably increase observers' reliance on stereoscopic slant cues, as compared with the sparse random-dot displays that have been commonly employed to study human perception of shape from binocular disparity and motion.