Abstract
The task of locating a target pattern within a scene is ubiquitous in vision. Past work on visual search typically used synthetic stimuli such as blobs and lines and has focused on a few low-level saliency cues. We build upon this work by investigating visual search for real-world objects. In particular, we examined human search efficiency by using an eye-tracker to record subjects' scan paths as they looked for designated targets in 2D arrays of randomly chosen natural objects.
In order to make principled comparisons across subjects and stimuli, we have developed a metric to quantify ‘search efficiency’. Using this metric, our experimental data indicate that while natural object search is typically not a parallel one, observers are considerably more efficient than an item-by-item serial search strategy. Matching of attributes, such as coarse color and orientation histograms in the periphery, drives the search, in stages, towards the target. Different attributes appear to be effective at different degrees of eccentricity. This suggests a computational approach to visual search that involves the creation of coarse maps of target likelihood in peripheral vision that are defined over simple attributes such as color and orientation. These maps then guide the next fixation in an iterative manner until the target has been precisely located. In addition to helping understand search by normal observers, this model makes interesting predictions about how search efficiency is likely to degrade in tunnel vision patients - patients with normal foveal vision but varying degrees of peripheral visual field loss.