Modeling Inefficiencies in Visual Search

Abstract

We have previously shown that visual search performance is well predicted by peripheral discriminability between patches containing multiple distractors and those containing distractors plus a target (Rosenholtz et al, in submission), which in turn is well predicted by a model in which the representation in peripheral vision consists of a rich set of summary statistics (Rosenholtz et al, in submission, and VSS 2009). Furthermore, we have developed an Ideal Saccadic Targeter model (saccades are made to the most likely target given the observations) to quantitatively predict mean number of fixations to find a target (Rosenholtz et al, VSS 2010). In this study, we examine four possible sources of inefficiency in human searchers relative to the ideal observer. (1) Memory. We compared the performance of the model with memory of the past k fixations against that of a memoryless model. (2) Location Uncertainty. We introduced location uncertainty by removing direct information about item location, so the model has to infer possible target locations from observations. The model has found the target if it fixates within one degree of it. (3) Saccade Length. Because humans prefer to saccade smaller distances (Zelinsky, 2008), we tested three ways of imposing this inefficiency: (a) strictly limiting saccade length, (b) having the model make intermediate saccades to a desired fixation point beyond a saccade limit, and (c) imposing a cost which depends on saccade length. (4) Ideal Decision Rule. We investigated the difference in search efficiency when using the optimal (maximum-a-posteriori) target location as opposed to picking the location with the highest targetness observation. Whether and how much memory was required depended on how saccade limitations were imposed. In general, the interaction between inefficiencies are complex. We show how the model predictions varies with these inefficiencies, and what sets of parameters best match observed human subject performances. Meeting abstract presented at VSS 2012