Abstract
Search is a fundamental and ubiquitous visual behavior. Here, we aim to model fixation search under naturalistic conditions and develop a strong test for comparing observer models. Previous work has identified the entropy limit minimization (ELM) observer as an optimal fixation selection model.1 The ELM observer selects fixations that maximally reduce uncertainty about the location of the target. However, this rule is optimal only if the detectability of the target falls off in the same way for every possible fixation (e.g., as in a uniform noise field). Most natural scenes do not satisfy this assumption; they are highly non-stationary. By combining empirical measurements of target detectability with a simple mathematical analysis, we arrive at a generalized ELM rule (nELM) that is optimal for non-stationary backgrounds. Then, we used the nELM rule to generate search time predictions for Gaussian blob targets embedded in hundreds of natural images. We also simulated a maximum a posteriori (MAP) observer, which is a common model in the search literature. To examine which model is more similar to human performance, we developed a double-dissociation search paradigm, selecting pairs of target locations where the nELM and the MAP observer made opposite predictions regarding search speed. By comparing the difference in human search times for each pair with the different model predictions, we can determine which model predictions are more similar to human behavior. Preliminary data from two observers show that human observers behave more like the nELM than the MAP. We conclude that the nELM observer is a useful normative model of fixation search and appears to be a good model of human search in natural scenes. Additionally, the proposed double-dissociation paradigm provides as a strong test for comparing competing models. 1Najemnik, J. & Geisler W.S. (2009) Vis. Res., 49, 1286-1294.
Meeting abstract presented at VSS 2015