Abstract
Many models of eye movements during search maximize the probability of saccadic targeting by selecting the location with the highest posterior probability of containing the target (Beutter et al., 2003; Torralba, 2003, Journal of the Optical Society of America A). Recently, a different model was proposed that uses knowledge of the varying visibility of the target/distractors across the retina to optimally plan sequences of saccades maximizing localization performance (optimal searcher, Najemnik & Geisler, 2005, Nature). Here, we compare the two models' distributions of 1st saccade endpoints to those of three human observers during a search for a bright Gaussian target among four dim distractors (Caspi et al., 2004). The contrast of the target as well as the distractors was varied over time, every 25 ms, by adding samples of random Gaussian noise (single frame SNR = 1.4). Display elements were placed equidistant around the circumference of a circle with a radius of 6.4 deg. The 1st saccade endpoints of human observers and the ideal saccadic targeting model clustered around the possible target locations. The pattern of 1st saccade endpoints of the ideal searcher varied with signal contrast, becoming more like the ideal target saccadic targeting model at low SNRs, and clustering around the center of the display for high SNRs. We present a quantitative approach to compare human results and model predictions by estimating the likelihood of observing the human saccade endpoints given that a particular underlying model is driving the eye movements.