Predicting Visual Saliency and Saccade Probability

Continually throughout the day, the brain must process incoming visual signals and transform them into appropriate actions, namely saccadic eye movements to the most behaviorally salient stimuli. The overarching problem that we address is how to use visual information to predict the parts of a scene that are most salient, and more specifically, to predict the targets of saccadic eye movements. Predicting the targets of saccades is a difficult problem for several reasons. First, there are many different cognitive influences on the saliency of a visual object. For example, a set of keys sitting on a desk might go overlooked when scanning the scene for a coffee mug, but will immediately draw the gaze of a viewer who has been locked out of the next room. Secondly, the saliency of a target is historyand state-dependent: a nearby object will often be targeted by a saccade over a more salient, but more distant, alternative. Finally, although repeated viewings of the same visual stimuli are necessary for data collection and proper analyses, saliency changes as objects become more familiar.