Abstract
Directing gaze toward different parts of a visual image enhances the visual processing of these selected regions. A straightforward hypothesis is that efficient saccadic eye movements select regions that maximize the information, or, equivalently, minimize entropy, relative to the image and the task at hand. However, in unconstrained visual exploration the definition of a functional utility function (the information that has to be maximized) is very complex. In addition, previous studies using well defined hard tasks (a visual search and a visual template-matching task respectively) have shown that the pattern of free fixation selections may maximize either the global, task-related information (Najemnik and Geisler 2005) or the local image-specific information (Renninger et al. 2007). To further elucidate the process underlying efficient fixation selection, we used a simplified experimental design. We recorded saccadic eye movements and perceptual performance in a visual 4AFC identification task in which large composite stimuli (i.e. non resolvable within one single fixation) were presented for 500ms, a duration which typically allows a maximum of two distinct fixation periods. Results suggest that subjects tend to adopt globally optimal strategies for fixation selection, minimizing the entropy related to stimulus identity. When several alternative strategies lead to entropy minimization, one scheme of fixation selection is maintained throughout the task (possibly the one minimizing motor effort). Furthermore, we observed that a bias in the probability of occurrence of specific stimuli did not affect oculomotor strategies except when it modified the spatial distribution of the maximally informative regions. Finally, we compared the perceptual performance in the free-gaze identification task with a model of strictly sequential static identification of local elements (as though at each fixation, only the element closest to gaze location was processed). Surprisingly, the performance in the free-gaze task was in general below that prediction. Possible explanations are discussed.
Meeting abstract presented at VSS 2013