Abstract
To evaluate the relative influence of target feature information and spatial context information in visual search, human eye movements were measured in a task where a small Gabor search target was randomly positioned at one of four pre-cued locations (the context) within a circular background of 1/f noise. All cued locations were 4 deg from the initial fixation point at the center of the display; the two outermost locations were fixed at the right and left of the display, 180 degrees apart, while the middle two locations were randomly chosen on each trial from 9 possible equally-spaced locations between the two outermost locations. Each trial began with the pre-cue display. In the “context-plus-features” condition, subjects maintained fixation until they initiated the onset of the search display with a button press, giving them the opportunity to extract target feature information before making a saccade. In the “context-only” condition, subjects also maintained fixation until the button press, but the search display did not appear until an eye movement was initiated. The duration of the search display was fixed at 500ms. Interestingly, the results suggest that in both conditions subjects planned two saccades prior to initiation of the first saccade, and thus in this task the eye movement strategy was largely determined by the context. The human eye movement patterns were compared with the predictions of several model searchers including the Bayesian ideal (Ideal), expected entropy minimization (EEM) and random (Ran). Both the Ideal and the EEM were found to predict the direction of human saccades relatively well; however, humans tended to fixate closer to the arc on which the targets lay than did the models.