Modeling Visual Attention in VR: Measuring the Accuracy of Predicted Scanpaths

Abstract

Dynamic human vision is an important contributing factor to the design of perceptually-based Virtual Reality. A common strategy relies on either an implicit assumption or explicit measurement of gaze direction. Given the spatial location of foveal vision, computational resources are directed at enhancing the foveated region in realtime. To obtain an explicit gaze measurement, an eye tracker may be used. In the absence of an eye tracker, a computational model of visual attention may be substituted to predict visually salient features. The fidelity of the resultant real-time system hinges on the agreement between predicted and actual regions foveated by the human. The contributions of this paper are the development and evaluation of a novel method for the comparison of human and artificial scanpaths recorded in VR. The novelty of the present approach is the application of previous accuracy measures to scanpath comparison in VR where analysis is complicated by head movements and dynamic imagery. An attentional model previously used for view-dependent enhancement of Virtual Reality is evaluated. Analysis shows that the correlation between human and artificial scanpaths is much lower than expected. Recommendations are made for improvements to the model to foster closer correspondence to human attentional patterns in VR.