Abstract
Two successful approaches to understanding lightness perception that have developed along largely independent paths are anchoring theory and Bayesian theories. Anchoring theory is a set of rules that successfully predict lightness percepts under a wide range of conditions (Gilchrist, 2006). Some of these rules are difficult to motivate, e.g., larger surfaces tend to look lighter than small surfaces. Bayesian theories rely on probabilistic assumptions about lighting and surfaces, and model percepts as rational inferences from these assumptions combined with sensory data. Here I reconcile these two approaches by showing that many rules of anchoring theory follow from simple, realistic assumptions about lighting and reflectance. I describe a Bayesian theory that makes the following assumptions. (1) Reflectances follow a broad, asymmetric normal distribution that is skewed towards low reflectances. (2) Lighting consists of multiplicative and additive components (Adelson, 2000). (3) The proportion of additive light tends to be low. These assumptions predict the main rules of anchoring theory, including: (a) The highest luminance in a scene usually looks white (anchoring to white), and (b) other luminances have lightnesses that are approximately proportional to luminance. (c) A perceived reflectance range of less than 30:1 is adjusted towards 30:1 (scale normalization). (d) When a low-luminance region becomes larger, its lightness increases, and the lightness of all other regions also increases (area rule). (e) The luminance threshold for glow increases with patch size. (f) Lightness percepts do not change when all luminances in an image are multiplied by a common scale factor. (g) Lightness constancy is better in scenes containing many distinct luminance patches (articulation). Thus anchoring theory can be formulated naturally in a Bayesian framework, and many seemingly idiosyncratic properties of lightness perception emerge as rational consequences of simple, realistic assumptions about lighting and reflectance.
Meeting abstract presented at VSS 2013