Abstract
Gauging the interest level of a person viewing a scene is potentially useful to the study of human behavior. While we know a lot about what attracts interest in an individual, e.gs. abrupt onsets, visual pop-outs, semantically incongruous scenes, we know less about what sustains interest. Here, we address this question in two sets of studies. In the first study, we measured interest sustainment to scenes of the same semantic category. Specifically, subjects viewed a series of scenes while we measured their viewing time, tracked their eyes, and asked them to rate their interest in the scene. In six separate sessions on separate days, subjects (n=17) viewed a series of 25 scenes of the same general category (aerials/cityscapes/indoors/landscapes/people, and a combined mix) in a self-paced manner. Subjects spent significantly longer time viewing landscapes than cityscapes, indoors, people, or the combined mix. Viewing time declined with trial number as expected but the longer mean viewing times on landscapes remained. The number of saccades/fixations the subject made while viewing the scene correlated significantly with viewing time, whereas saccade and fixation durations did not. A second experiment on new subjects (n=48) with new (75/category) scenes confirmed overall significantly longer viewing times on landscapes versus people; subjective interest ratings (landscapes>combined>people) aligned with viewing times. In the second study, we present two scenes from different categories side by side and compare the time subjects spend viewing one scene versus another and ask subjects to provide a subjective interest rating for each of the scenes in the pair. The study is ongoing. In both studies, we correlate viewing time and interest rating with measures of image complexity based on fractal dimension and Jansen-Shannon divergence in order to test the hypothesis that the complexity of a natural scene is a good proxy for sustained interest.
Meeting abstract presented at VSS 2018