A common characteristic of psychophysical eye-tracking data taken during free-viewing conditions is a strong central bias in the set of fixation locations. Although trends in photographic composition may contribute to a centralized distribution of fixations, a more fundamental explanation arises from a train of explicit overt fixations. We show first that a 2D random walk of fixations implicitly encodes a central bias. Based on that result, we implement a saccade controller for overt fixational control. The saccade controller consists of three distinct modules: a peripheral priority map (PPM), a fixational history map (FHM), and an history-biased priority map (HBPM). In the absence of top-down tuning, the PPM functions in the same manner as a saliency map, while the FHM provides inhibition of return functionality. The HBPM combines input from the central candidate for attention, the PPM, and the FHM to determine the next fixation target. Using the saccade controller in conjunction with the AIM saliency algorithm (which does not contain a central bias component) we demonstrate an improved performance in reproducing more human-like fixation patterns.
Meeting abstract presented at VSS 2013