Abstract
The question of where we look within a visual scene has long been studied by monitoring fixation positions and their relationship to low-level (color, intensity, orientation, flicker and motion) and higher level (object-based, global depth or contextual) visual features in the image. Besides fixation location, the study of other saccadic parameters, especially saccade direction, can give new insight into (the preattentive mechanism of) overt attention. For simple slanted grid planes, spontaneous saccades follow the orientation of surface slant and tilt (a depth cue combination, as in conscious perception). This strategy may extend to complex scenes with natural stimuli. We test this assertion for two types of scene contexts (natural forest vs man-made urban). We examined eye movements and saccade direction in 16 observers while they free-viewed static color images over 7s. Observers' saccades were principally vertically and horizontally oriented for both conditions, but of importance, were more horizontally (p<0.05) and less vertically (p<0.05) oriented for urban than for forest context. In particular, the first saccades tended to follow the main horizontal and vertical directions of the scene and its ground surface orientation in depth: the slant. The saccade direction mechanism appears to be linked to either some low-level gradient or edge orientation, higher-level convergence of vanishing lines, or a more global gist contextual feature computation. This fast saccadic programming might rely on a visual mechanism to infer scene spatial layout: surface orientation, perspective, openness and scene category, in the absence of vestibular information. This result argues in favor of a feed-forward high-level scene representation, which could be accessed in parallel during the deployment of overt attention.
Meeting abstract presented at VSS 2014