Abstract
Previous research has shown that the contextual information in real-world scenes helps guide visual attention when searching for a target within the scene (Torralba et al., 2006). However, it is unknown whether such contextual guidance can occur in the absence of semantic scene recognition. To address this question, we generated texture patterns that were unrecognizable as real-world scenes, yet preserved the statistical regularities of real-world scenes (i.e., the global pattern of orientation and spatial frequency information). In each texture, we imbedded the image of a pedestrian at a location where a pedestrian was likely to appear with either a low probability or a high probability (based on an independent set of rankings using the original, real-world images). On each trial, observers were instructed to locate the pedestrian and indicate, as quickly and accurately as possible, the direction in which the pedestrian was facing. Response times for the high probability trials (M = 1989 ms) were reliably faster than the low probability trials (M = 2593 ms) (t(8) = 7.09, p <.001). This difference could not be explained by differences in absolute screen position: the low-probability and high-probability locations were matched across images (i.e., the low-probability location for one image was the high-probability location for another). This difference also could not be explained by differences in local visibility, because a control experiment showed that when the background is erased, except for a local window around the pedestrian, there was no difference in reaction time for high and low-probability locations (p > .05). Thus, the advantage for the high-probability locations arises from the global context of a particular image. Combined, these results suggest that the statistical regularities of real-world scenes can guide the deployment of visual attention, even in the absence of semantic scene recognition.