Purchase this article with an account.
David Brainard; Bayesian models of object color perception. Journal of Vision 2009;9(14):6. doi: 10.1167/9.14.6.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
An element common to many aspects of perception is that the image information available at the retina is insufficient to uniquely determine the physical configuration of the scene. To provide useful perceptual representations of the environment, the visual system must combine the data available in the retinal image with additional constraints. Understanding how to conceive of and model this process is a long-standing and central problem in our field. This general problem is encapsulated by the model system provided by the special case of object color perception. Object color is a perceptual correlate of object surface reflectance, and here the ambiguity arises because the light reflected from objects depends both on their surface reflectance and on the illumination. A visual system that delivers a percept of object color that is independent of the illumination (and of variation in other scene variables) is referred to as color constant. It is now generally agreed that the human visual system exhibits a considerable degree of color constancy, but that this constancy is not perfect. In this talk, I will review Bayesian models of human color constancy. These models start by formulating an explicit estimation problem: how can object surface reflectance be estimated from the retinal image? The Bayesian framework provides a natural way to answer this question, as it allows for expression of how the image data constrain the estimate as well as a mechanism for incorporating constraints provided by the statistical structure of the natural environment. In addition to showing specific cases where Bayesian models account for human color perception, I will also attempt to make a number of broader points. These may include a) the importance of formulating explicit linking hypotheses that connect a model's estimates with psychophysical data, b) the desirability of seeking modeling principles that generalize naturally, c) the fact that Bayesian methods are largely silent about the nature of the neural mechanisms that implement them, d) that this is OK, and e) that although Bayesian estimation is optimal in a well-defined statistical sense, Bayesian models are none-the-less capable of accounting for both success and failures of color constancy.
This PDF is available to Subscribers Only