Rapid Saliency Identification for Selectively Rendering High Fidelity Graphics

Images rendered using global illumination algorithms are considered amongst the most realistic in 3D computer graphics. However this high fidelity comes at significant computational expense. It has long been a goal of computer graphics to create realistic images at an interactive rate, unfortunately the computational cost of these algorithms prohibits this. As a direct result of this, the realism of graphics generally decreases as the synthesis of images approaches real time. In this thesis we examine the concept of realism as it relates to computer graphics. By looking at how we, as humans, perceive the world around us and printed, or projected versions of that world, we can gain useful insight into the requirements of computer graphics. We present methods for reducing the time required for the production of rendered images based on human perception. By studying which areas of an image are considered important and which areas go unobserved it is possible to direct computational effort only where it is needed. Our approach is based on accepted psychophysical data. Although previous methods have used models of human visual perception to control rendering algorithms, our work takes a novel approach. By using prior knowledge of scene content and the efficient use of modern graphics hardware, we interactively create a map of required pixel quality. This map can then be used to reduce the complexity of a rendering algorithm on a per pixel level. This is the opposite of most previous approaches which attempt instead to save time by stopping the rendering calculation when pixels reach a certain quality threshold. Additionally we present a new method in which a combination of the GPU and the CPU can be used together to selectively render images.