Abstract
Previously (Rosenholtz, Vis. Research, 1999; Perception & Psychophysics, 2001), we have presented the Statistical Saliency Model for visual search. This model says that an item in a display is salient if its feature vector is an outlier to the local distribution of feature vectors, according to a parametric statistical test for outliers. In particular, saliency is given by essentially the number of standard deviations between a given feature vector and the local mean. A simple model — that visual search is easier the greater the saliency of the target — has been shown to qualitatively predict the results of a number of search experiments involving low level features such as color, motion, and orientation.
We will present a computational form of the Statistical Saliency Model, which operates on arbitrary images, and consists of biologically inspired mechanisms. As with preattentive texture segmentation (Rosenholtz, Proc. ECCV, 2000) there need not be a dichotomy between models that are statistically inspired and those that are biologically inspired, nor between models based upon the desired function of a saliency computation, as opposed to its implementation in neural hardware. Where statistical models of perceptual phenomena are appropriate, as, we argue, in visual search, deriving models based upon what the brain tries to do, and why, as opposed to how it might do it, can lead to fewer free parameters, with more intuitive interpretations, and easier design of experiments to determine those parameters. The Statistical Saliency Model was designed with an eye towards the brain's purpose (detect outliers or unusual items) as opposed to its possible neurocomputational mechanisms, yet its statistical test may be implemented using such mechanisms.