Generating Sequence of Eye Fixations Using Decision-Theoretic Attention Model

Human eyes scan images with serial eye fixations. We propose a novel attention selectivity model for the automatic generation of eye fixations on 2D static scenes. An activation map was first computed by extracting primary visual features and detecting meaningful objects from the scene. An adaptable retinal filter was applied on this map to generate "Regions of Interest" (ROIs), whose locations corresponded to those of activation peaks and whose sizes were estimated by an iterative adjustment algorithm. The focus of attention was moved serially over the detected ROIs by a decision-theoretic mechanism. The generated sequence of eye fixations was determined from the perceptual benefit function based on perceptual costs and rewards, while the time distribution of different ROIs was estimated by a memory learning and decaying model. Finally, to demonstrate the effectiveness of the proposed attention model, the gaze tracking results of different human subjects and the simulated eye fixation shifting were compared.