We present and validate a region-of-interest (ROI) rendering system, in which the rendering parameters of the image are adapted based on models of human visual attention. Rendering realistic computer graphics imagery using global illumination algorithms is computationally expensive and this process can take many hours. The inclusion of participating media, such as fog or dust, considerably compounds this problem, since the scattering of the light due to the medium must also be computed. However, human perception of the final rendered image is not fully dependent on all of these computations. By removing unnecessary details, this performance overhead can be substantially reduced. Perceptually-based selective rendering algorithms achieve this by exploiting limitations of the human visual system, and the rendering quality is varied spatially across the image. In areas of high perceptual importance the algorithm adapts to provide a more accurate solution relative to less significant areas. Two general visual attention processes exist and these are referred to as bottom-up and top-down. Computational models of the bottom-up process indicate where a human observera??s attention would be automatically drawn while free-viewing images. Features that attract attention are referred to as salient. Models of top-down processes are related to the visual task being performed by the observer. Our ROI rendering system accounts for both bottom-up and top-down computational models by introducing a novel importance map (IM). This IM is a combination of a saliency map and a task map. The saliency map is based on low-level features such as colour, intensity, and orientation. The task map is based on the objects related to the visual task. This thesis assesses the perceived quality of selectively rendered still images and animations using IMs. Results of qualitative psychophysical experiments, using human observers, are presented as well as quantitative differences metrics. The results demonstrate that a perceived high quality can still be achieved while computation is reduced by up to an order of magnitude.