Abstract
Perception of three dimensional (3D) shape is typically biased. Bayesian models of perception attribute such biases to the influence of prior knowledge that acts to restrict perception to likely 3D shapes. However, an infinite set of Bayesian models can predict the same perceived shape, for even the simple situation of a single cue and associated prior. By further constraining a Bayesian model to predict observer variance for a task, one can understand more details of the cue and prior information (i.e. the shapes of the likelihood functions and prior distributions).
We demonstrate a method of calculating the prior distributions and likelihood functions of a Bayesian model by considering settings made by observers in a 3D shape matching experiment. Stimuli contained an outline cue, a shading cue, or both. Observer estimates show a large bias towards frontoparallel surface orientation when using either cue, or a combination of both. A Bayesian cue combination model, constructed using Gaussian likelihood functions and prior distributions, can predict mean shape settings when observation variance is not considered. However, when the model is constrained to fit observation variance as well as mean shape settings, prior distributions that either depend on the stimulus or have heavy tails are needed.
We conclude that strong priors affect observers’ perception of shape in our stimuli, and that these priors must be represented by distributions that are heavier tailed than Gaussians, similar to priors that have been shown to explain biases in speed perception.
Meeting abstract presented at VSS 2012