Abstract
Search performance with synthetic images improves when a target appears at an expected location (indicated by an artificial cue) rather than an unexpected (uncued) location. Similarly in real scenes the accuracy of the 1st saccade improves when the target appears at an expected location vs. an unexpected location (Drescher et al., 2003). The attentional processes mediating the context effect in real scenes are unknown (Chun, 2000). One standard explanation of the classical cue effect is that the observer deploys limited attentional resources to the cued location improving the quality of processing. A competing theory is that the benefit occurs due to differential weighting of visual information (for linear weighting: Kinchla et al., 1995; for Bayesian priors: Eckstein et al., 2002; Shimozaki et al., 2003; Oliva et al., 2003). Here, we measured the accuracy of the 1st saccade during search in 24 real scenes when the target was placed at expected and unexpected locations. The study controlled for target detectability and initial retinal eccentricity. A third condition in which the target was absent from the images allows to distinguish between the limited resources vs. differential weighting models. Human results were compared to computational implementations of the two attention models. Results: Accuracy of the 1st saccade averaged across observers improved by 3.6 deg (p < 0.05) when the target was at an expected location vs. an unexpected location. Observers' 1st saccade endpoints in target absent images were significantly closer to the expected rather than the unexpected locations, consistent with the differential weighting model (Bayesian priors) and inconsistent with the limited resources model. Conclusions: The results suggest that the increased saccadic accuracy in real scenes when the target appears at an expected (rather than unexpected) location is due to a higher observer weighting (prior) of visual information at likely target locations.