Abstract

Recent computer models of active vision adopt the broad categories of stimulus-driven and concept-driven eye movements to control the saccade-like movements of their cameras. Although low-level target selection may be accomplished by assigning priorities to areas on a saliency map, the question of how high-level feedback might influence saccadic movements is open to debate. One widely recognized solution to this problem defines this top-down input in terms of a hypothesistesting strategy. According to this model, foveal information and peripheral information are combined to suggest various hypothetical descriptions of a scene. Saccadic movements would then be used to confirm or reject the most likely of these hypotheses. An experiment was conducted to determine under what conditions a hypothesis-testing scheme is used and at what stage such information becomes available to the human occulomotor system. Subjects were asked to discriminate among four similar patterns while their eye movements were recorded by a two-dimensional binocular eye-tracker. The initial saccade made to each pattern was then used to determine if and when targets providing disambiguating information were preferred over targets providing only redundant information. This paradigm permits a systematic study of these top-down influences on the planning of saccades and may have implications for models of active vision.

© 1990 Optical Society of America