Visual attention using 2D & 3D displays

Abstract

In the past three decades, robotists and computer vision scientists, inspired by psychological and neurophysiological studies, have developed many computational models of attentions (CMAs) that mimic the behaviour of the human visual system in order to predict where humans will focus their attention. Most of CMA research has been focussing on the visual perception of images and videos displayed on 2D screens. There has recently, however, been a surge in devices that can display media in 3D and CMAs in this domain are becoming increasingly important. Research in this context is minimal, however. This thesis attempts to alleviate this problem. We explore the Graph-Based Visual Saliency algorithm [68] and extend it into 3D by developing a new depth incorporation method. We also propose a new online eye tracker calibration procedure that is more accurate and faster than standard processes and is also able to give confidence values associated with each eye position reading. Eye tracking data is used to evaluate CMAs. We use our novel eye tracking method to create a 2D/3D video eye tracking dataset obtained from 50 people. A statistical analysis is performed to locate where perception differs in 2D and 3D in videos. Taking advantage of the uncertainties associated with our eye tracking data, we also propose a novel Gaussian mixture model for computing eye tracking heat maps.