Do low-level visual features have a causal influence on gaze during dynamic scene viewing?
23.311, Saturday, May 11, 8:30 am - 12:30 pm, Royal Ballroom 6-8
Parag Mital1, Tim J. Smith2, Steven Luke3, John Henderson3; 1Department of Computing, Goldsmiths, University of London, 2Department of Psychological Sciences, Birkbeck, University of London, 3Department of Psychology, University of South Carolina, Institute for Mind and Brain
Recent evidence of eye-movement behavior during dynamic scene viewing has shown that low-level visual features such as motion are of greater magnitude and contrast at fixation than at non-fixated locations. However, it is still unclear if the features at fixation are causal in attracting attention or merely correlated with higher-level features such as people. We investigated the influence of motion, edge, and luminance contrast features on eye-movements by either retaining or filtering out each of these features in video clips while participants rated their preference for a 3 second clip. Do participants exhibit the same eye movements and attend to the same locations of a scene when individual or combinations of visual features are removed? Our results suggest that motion and edge information are required for normal eye movement behavior in dynamic scenes. Removing contrast has minimal impact on gaze behavior as measured by gaze entropy, fixation durations, and saccade amplitudes. Removing either edge or motion information significantly increases fixation durations and shortens saccade amplitudes. However, the removal of edge information on its own has a significantly greater impact on fixation durations and saccade amplitudes than the removal of motion. This difference may be due to the fact that removing motion does not impair the initial gaze bias towards human figures observed in the original clips whereas removing edges does. Edge information is required to locate people in a dynamic scene and create typical eye movement behavior. Motion is important for typical timing and amplitude of gaze shifts but not for locating typical focal features such as people. These results suggest that typical gaze behavior in dynamic scenes is due to an interaction between low-level visual features and scene semantics which can be used to compensate for impoverished low-level cues and guide attention to relevant scene content.