Speaker: Hiroyuki Sakai
The reproduction of real world light phenomena has always been a major goal in computer graphics. While the interaction of light and matter are well understood from a physical standpoint, the efficient simulation is still a challenge. Specific phenomena, such as direct illumination or direct shadows, can be rendered relatively easy in real-time by means of GPU rasterization. Indirect effects, where the intensity of a certain point not only depends on a particular light source but also on its surroundings, increase the computation complexity considerably.
Real-time rasterization approaches for rendering such effects generally require specifically tailored algorithms and the use of significantly simplified and approximative light transport models. Notable recent examples of such approaches are light propagation volumes and voxel cone tracing.
In contrast to such rasterization approaches, path tracing methods actually simulate the light transport by taking various light paths from the light source to the viewer as well as the interaction of light and matter into account. This inherently allows for a more accurate reproduction of real world light phenomena at the cost of computation time. To generate visually pleasing results, a sufficient number of paths has to be evaluated. A low number of paths decreases the computation time but leads to noise in the final rendering.
The major aim of this thesis is the denoising of such undersampled images with the recently developed adaptive manifold filtering algorithm to ultimately increase the quality of path traced images which were rendered with a low number of samples in a comparatively short amount of time. The reconstruction of specular (mirror-like) and glossy surfaces is of particular interest.