Abstract
Where we look in a real-world scene depends on a number of factors, including stimulus properties (e.g., local contrast), task, and constraints of human visual processing (e.g., acuity loss in peripheral vision). Previous work has examined the role of bottom-up saliency (Itti & Koch, 2000) and task-driven eye-movement planning (Najemnik & Geisler, 2005; Renninger et al., 2007) in determining eye movements. Purely bottom-up saliency cannot account for eye movements in real-world scenes; rather, top-down influences have to be considered (Oliva et al., 2003; Henderson et al., 2007; Tatler et al., 2011). Furthermore, humans seem to use eye-movement strategies that account for the limitations of peripheral vision and maximize task-specific information gain (Peterson & Eckstein, 2012). In this work, we test an alternative way of conceptualizing eye movements in terms of visual conspicuity (Engel, 1971). The conspicuity of a target measures the maximum eccentricity at which the target viewed peripherally looks the same as an identical target presented foveally. This measure includes both bottom-up and top-down aspects of visual recognition and is constrained by the limitations of peripheral vision. Target conspicuity is correlated with the probability of finding a target in simple displays (Engel, 1977) and with search times in restricted real-world settings (Toet et al., 1998). We extend these findings to a "free-viewing" memory task and unrestricted real-world scenes. We use images from the MIT Eye Tracking Dataset (Judd et al., 2009) and show that the regions that elicit fixations have lower conspicuity (i.e., are harder to resolve in the periphery) than the regions that do not elicit fixations. In addition, conspicuity succeeds where a state-of-the-art saliency model (Judd et al., 2009) does not; regions incorrectly predicted to be salient have higher conspicuity values than correctly predicted regions.
Meeting abstract presented at VSS 2013