Abstract
Visual search is a fundamental behavior, yet little is known about search in natural scenes. Previously, we introduced the ELM (entropy limit minimization) fixation selection rule, which selects fixations that maximally reduce uncertainty about the location of the target. This rule closely approximates the Bayesian optimal decision rule, but is simpler computationally, making the ELM rule a useful benchmark for characterizing human performance. Previously, we found that the ELM rule predicts several aspects of fixation selection in naturalistic (1/f) noise, including the distributions of fixation location, saccade magnitude, and saccade direction. However, the ELM rule is only optimal when the detectability of the target (the visibility map) falls off from the point of fixation in the same way for all potential fixation locations, which holds for backgrounds with relatively constant spatial structure, like statistically stationary 1/f noise. Most natural scenes do not satisfy this assumption; they are highly non-stationary. By combining empirical measurements of target detectability in natural backgrounds with a straight-forward mathematical analysis, we arrive at a generalized ELM rule (nELM rule) that is optimal for non-stationary backgrounds. The nELM searcher divides (normalizes) the current target probability map (posterior-probability map) by the estimated local contrast at each location in the map. It then blurs (convolves) this normalized map with the visibility map for a uniform background. The peak of the blurred map is the optimal location for the next fixation. We will describe the predictions and performance of the nELM searcher.
Meeting abstract presented at VSS 2014