Abstract
In previous work we demonstrated that the search for a categorical target depends on the hierarchical level in which the target is cued. Targets cued at the subordinate level displayed an advantage in measures of guidance while basic-level cues were verified the fastest. The Category-Consistent Feature (CCF) model extracts the important features of a target category, computationally-defined as high-frequency and low-variability features across the category exemplars. This model predicted behavioral trends across hierarchical levels, but can it also predict search behavior on individual trials? Participants (n=26) searched 6-item displays for a text-cued target in subordinate (n=48), basic (n=16), and superordinate-level (n=4) conditions (68 categories in total). Targets and distractors were images of objects cropped from ImageNet and Google. The CCF model represents a category by averaging SIFT+color bag-of-words feature histograms of exemplars and using a signal-to-noise ratio to prune away the less informative features in the averaged histogram. The resulting category-specific CCF histogram is then compared to the raw feature histograms extracted from the target and distractors in the search display to derive a categorical priority map, which we use to predict behavior. Correlating these independent model predictions with search performance, we found that the CCF model successfully predicted both the time taken to initially fixate the target (target guidance) and the time needed to verify the target as a member of the cued category after its initial fixation (target verification). Importantly, we obtained these results in the context of a hierarchical structure spanning three levels, 68 categories, and 4,800 image exemplars, thereby demonstrating the robustness of our generative modeling approach. By using the CCF model to represent the within-category feature structure of common objects, it is now possible to predict categorical search behavior, not just in the aggregate, but on individual trials.
Meeting abstract presented at VSS 2016