Selective rendering with graphical saliency model

Abstract

In this work, we firstly identify the shortcomings of the existing work of selective image rendering. In order to remedy the identified problems, we put forward the concept and formulation of a graphical saliency model (GSM) for selective image rendering applications, in which the sampling rate is determined adaptively according to the resultant saliency map under a computation budget. Different from the existing visual attention (VA) models which have been devised for natural image/video processing and applied to image rendering, the GSM considers the characteristics of the rendering process and aims to detect regions which require high computation to be rendered for good use of the said budget. The proposed GSM improves a VA model by incorporating a metric of rendering complexity. Experiment results show that, under a limited computation budget, selective rendering guided by the proposed GSM can achieve better perceived graphic quality, compared with that merely based upon a VA model.