Abstract
The interaction of top-down and bottom-up control of visual attention is of central importance for our understanding of vision, and most of its extensive study has employed the paradigm of visual search. However, little is known about the dynamics of top-down and bottom-up mechanisms during demanding search tasks in complex scenes that guide our attention so efficiently in everyday situations. Here, we present and apply a novel method to estimate the time course of visual span, that is, the area around gaze fixations from where visual features can exert top-down or bottom-up control of attention. The method assumes that a larger visual span allows larger areas of relevant information to attract eye movements for its inspection within a single, central fixation (center-of-gravity effect, Findlay, 1982). Indeed, the distribution of gaze fixations can be predicted by convolving the distribution of relevant information with a Gaussian kernel whose size matches the visual span (Area Activation Model, Pomplun et al., 2003). In this study, we computed separate top-down (Hwang et al., 2009) and bottom-up saliency maps (Itti & Koch, 2001) for 160 real-world search displays and convolved each of them with Gaussian kernels of different sizes. Those sizes that resulted in the best predictors of the positions of 15 subjects' search fixations were taken as estimates of visual span. We used this method to estimate the strength and visual span of top-down and bottom-up control of attention by display features during different phases of the search process. Top-down control was found to be weak initially but to quickly dominate search while narrowing its focus, whereas bottom-up control revealed slowly diminishing influence and a constantly large visual span. These results suggest that, throughout the search process, accumulating scene knowledge determines the dynamics of attentional control, which is not reflected in current models.
Grant Number R15EY017988 from the National Eye Institute to Marc Pomplun.