Abstract
Models of visual search typically input a target and a distribution of distractors (along with other parameters), and output a prediction of search ease, e.g. reaction time, percent correct performance, or some qualitative measure. Models can also input only the distractor distribution, and output a threshold, i.e. a prediction of the target satisfying some minimal requirements for search ease. For example, in search for a target of a unique size a model might report the minimum target size required to achieve a certain percent correct performance at the search task.
In N-D feature spaces, the concept of a threshold generalizes to search isocontours — a set of locations in feature space representing a set of targets, each satisfying the desired requirements for search ease. E.G. for a given set of distractors, a search isocontour might indicate the set of all targets yielding a certain percent correct performance. Isocontours can be predicted by a model, or determined empirically.
Search isocontours give us, at a glance, a more complete image of the relationship between the target, distractors, and search ease. Predicted isocontours elucidate differences between models, and identify key experiments for choosing between those models. Isocontours offer a new pictorial way of seeing the effects of set size, distractor heterogeneity, and other factors.
Saliency Model (Rosenholtz, Vision Research, 1999; Perception & Psychophysics, 2001) isocontours are just ellipses representing the covariance of the distractors. (Rosenholtz, J. Exp. Psychology, 2001) demonstrated how to make predictions of signal detection theory models for arbitrary distributions of targets & distractors, and for N-D feature spaces. Based on this, one can determine search isocontours for SDT-based models. I will present search isocontours for several models and a number of examples. In addition, I will discuss issues in efficiently finding search isocontours empirically.