Abstract
In visual search, humans use eye movements to direct the fovea at potential target locations in the environment. Do humans employ rational eye movement strategies while searching for targets in cluttered environments? To answer this question, we derived the Bayesian ideal visual searcher for tasks where a known target is placed at an unknown location within a background of 1/f noise. We constrained the ideal searcher to have the same falloff in target detectability with eccentricity as humans. We find that humans achieve near-optimal performance in this search task, suggesting that humans must be selecting their fixation locations efficiently. To explore this further, we compared eye movement statistics of humans, ideal searchers, and suboptimal searchers that do not select fixation locations optimally but still integrate information perfectly across fixations. Remarkably, human search patterns match those of the parameter-free ideal searcher for most of the statistics we have examined, including: (1) the spatial distribution of fixation location, (2) the distribution of saccade lengths, (3) the change of mean distance of fixations from the center of the search area as search progresses, and (4) search time as a function of target/noise contrast and target position. A particularly interesting suboptimal searcher is the MAP searcher (which always fixates the most likely target location) because the MAP fixation strategy is the basis for most existing models of eye movements in visual search. Although the MAP searcher shares many eye movement statistics with humans and ideal, and achieves near-optimal performance, it can be rejected as a model of human search because it distributes fixations across the search area in a spatial pattern that differs from human and ideal. Also, humans substantially outperform suboptimal searchers that select fixation locations at random (with or without replacement), allowing us to conclusively reject all possible random search models.
Supported by NIH grant R01EY02688.