Selective image rendering based on an improved perceptual saliency model

Abstract

Computer graphics is becoming more and more important in film and game industries. However, photorealistic image rendering is still very time consuming. To reduce the rendering cost has always been a difficult task for researchers. Visual attention guided rendering is one of the most important areas of selective rendering. This project is to develop a perceptual saliency model and implement it as a selective rendering scheme which generates better quality images compared to an existing visual attention based rendering with the same amount of rendering cost. Our implementation is based on Physically Based Rendering Tool (PBRT). A blended saliency map which contains visual attention information and local sample contrast information is proposed and it guides the sampling process of the selective rending scheme. Integrating local sample contrast information into visual attention rendering in our experiment has demonstrated a better result than pure visual attention based rendering.