VISUAL ATTENTION IN DYNAMIC NATURAL SCENES

Abstract

The focus of this PhD work has been to gain a better understanding of human visual attention and eye movements while viewing dynamic natural scenes (e.g. videos and movies). This thesis is arranged in two parts. The first section describes the results of some eye movement experiments that investigated the effects of scene transitions in videos on fixation behaviour. Our results showed that scene transitions appeared to lengthen the durations of on-going fixations and decrease the durations of follow-up fixations. In addition, if the contrast at the fixated location in the new scene was sufficiently high, the visual system appeared to prefer continuing the fixation rather than pick a new location. The second section describes a novel computational model of saliency that makes use of gist information to compute the final saliency map. The model was tested with eye movement data from 32 subjects, and demonstrated comparable performance to several well-known models of visual saliency.