Abstract

V. Navalpakkam, L. Itti, Modeling the influence of target and distractor knowledge on visual search, In: Proc. Society for Neuroscience Annual Meeting (SFN'04), Oct 2004. (Cited by 7)

Abstract: Previous research on visual search suggests that knowledge of the target leads to a multiplicative increase in the activity of neurons that encode target features (i.e., gain factor > 1). We investigate how the added knowledge of the distractors influences the optimal choice of gain factors on target and non-target features, such that the target salience increases relative to the distractors. Assuming Poisson firing for neurons tuned to various visual features, we wish to decide the gain factor on each feature type that will maximize target detection and minimize distractor detection. If a distractor has the same amount of a feature as the target, the distributions of responses from neurons tuned to that feature fully overlap, and target and distractors will be detected equally. If a distractor has a greater amount of a feature than the target, then the distribution of distractor responses is shifted towards higher values, and a negative gain factor will optimize detection. Further, if the distractor has a new feature missing from the target, a negative gain factor on the new feature will optimize detection. To verify these hypotheses, we designed search arrays with one target and equal numbers of distractors of type SAME (same amount of feature as in target), MORE (greater amount of feature), and NEW (new feature missing from target). We ran 7 subjects for 600 trials each for 5 days and recorded their eye movements. Measurement of the number of fixations on each type of distractors revealed significantly more fixations on SAME distractors than MORE (p < 10^-10) and NEW (p < 10^-8). Our results suggest that subjects use a negative gain factor on all features present less in target than distractors. This prediction of negative gain at the behavioral level suggests the presence of neurons encoding the absence of features, whose activity is negatively correlated with those detecting the presence of features, leading to interesting predictions for search asymmetry.

Themes: Computational Modeling, Human Psychophysics, Model of Bottom-Up Saliency-Based Visual Attention, Scene Understanding