Purchase this article with an account.
Pascal Mamassian, Ross Goutcher; A Bayesian Model of Structure-from-Motion Perception. Journal of Vision 2004;4(8):77. doi: 10.1167/4.8.77.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
The structure of a rotating object can be inferred from the relative trajectory of its features. In particular, the slant of a rotating plane can in theory be recovered when it is seen under perspective projection. We are here interested in the biases in perceived slant, an issue that is essential to fully understand how human observers use this structure-from-motion cue. Our stimuli consisted of a horizontally-oriented convex wedge rotating about a vertical axis. The object was defined by small texture dots whose positions were uniformly distributed in the image. The inner angle of the wedge was 90 degrees whilst the overall orientation of the wedge varied about the horizontal axis from trial to trial. The task of the observer was to decide whether the top or bottom plane of the wedge was more slanted. The vertical position of the wedge was randomized to remove any confounding speed cue. Individual observers' data revealed significant biases to see the top plane as less slanted than the bottom one. In other words, floor surfaces appeared less slanted than they should, and ceiling surfaces more slanted. These data cannot be explained by the often reported regression to the fronto-parallel plane of the image. Instead, our results reveal slant preferences that affect observers' judgments of structure-from-motion. We have modeled our results with a simple Bayesian model that combines two major components. The likelihood represents knowledge of the trajectory of dots under perspective projection for given object slants. The prior probability reflects intrinsic preferences for particular slants. This Bayesian model is then used to infer the prior probability on slant that best predicts the observers' data.
This PDF is available to Subscribers Only