Abstract
The mechanisms of visual attention have been well studied in controlled experimental paradigms with simple stimuli. Experimental evidence suggests that both goal-directed and stimulus-driven mechanisms can guide attention. However, the way in which attention is guided by observers when viewing complex and dynamic natural scenes has only recently begun experimental evaluation. In previous research, we have shown that attention, as evidenced by eye movements, is stimulus-driven when observers freeview static natural scenes (Parkhurst, Law & Niebur, Vis Res 2002). Furthermore, using a computational model of attention, the stimulus-dependence of attention was found to be greatest just after stimulus onset. This work left open the question of the degree to which motion guides attention in dynamic natural scenes. To address this question, we implemented a computational model of visual selective attention that calculates stimulus salience based on color, intensity, orientation, and motion. To accomplish this in a biologically realistic way, neural mechanisms of visual processing thought to occur in the primate visual system were incorporated into the model. Each feature is processed at a series of spatial scales using realistic temporal parameters derived from neurophysiological data. The color and intensity feature pathways use temporal properties derived from the parvocellular and magnocellular pathways, respectively. Center-surround receptive field representation and scale normalization are implemented through divisive inhibition. Given a dynamic natural scene, the model produces a dynamic saliency map that indicates the most salient spatial locations in the scene over time. By quantitatively comparing the ability of the model to predict eye movements made by human observers with and without motion processing enabled, the degree to which attention is guided by motion in dynamic natural scenes can be determined.
Support: an NIMH NRSA predoctoral fellowship to DP and an NSF CAREER grant to EN