Perceptually Driven Radiosity
Despite its popularity among researchers the radiosity method still suffers some disadvantage over other global illumination methods. Usual implementations of the radiosity method use criteria based on radiometric values to drive the computation to decide about sufficient mesh quality or to estimate the error of the simulation process and to decide when the simulation can be safely terminated. This is absolutely correct for the case of radiometric simulation, when the user is interested in actual values of radiometric quantities. On the other hand, the radiosity method is very often used just to generate pictures for the human observer and those pictures are not required to be the results of correct physical simulations, they just have to look the same. The results of research on human visual performance and visual signal processing can be built into the image synthesis algorithm itself under some circumstances and guarantee that no effort will be spent on computing changes that are only marginally important for the human observer. In the area of image processing, perceptual error metrics are used for image comparison and image coding that enable to better predict the differences between two images as opposed to the perceptually inappropriate and widely used meansquared error metrics. Tone reproduction operators known from image synthesis make it possible to map a bright scale of image luminances onto a narrow scale of CRT luminances in such a way that the perceived CRT image produces the same mental image as the original image. Perceptually-driven radiosity algorithms exist, which use various methods to control the optimum density of the finite-element mesh defining the scene that is being rendered, to include only visible discontinuity lines into this mesh, and to predict the convergence of the method. We will describe an hierarchical extension to the Monte Carlo radiosity that keeps the accuracy of the solution high only in the area immediately visible from the point of observation. We will also present a comparison of different perceptual and radiometric termination criteria for a view-independent version of Monte Carlo radiosity.