Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Filtering techniques for low-noise previews of interactive stochastic ray tracing

: Schwenk, Karsten
: Fellner, Dieter W.; Dachsbacher, Carsten

Fulltext (PDF; )

Darmstadt, 2013, XIII, 186 pp.
Darmstadt, TU, Diss., 2013
URN: urn:nbn:de:tuda-tuprints-35906
Dissertation, Electronic Publication
Fraunhofer IGD ()
ray tracing; noise reduction; global illumination; interactive rendering

Progressive stochastic ray tracing algorithms are increasingly used in interactive applications such as design reviews and digital content creation. This dissertation contains three contributions to further advance this development.
The first contribution is a noise reduction method for stochastic ray tracing that is especially tailored to interactive progressive rendering. Highvariance light paths are accumulated in a separate buffer, which is filtered by a high-quality, edge-preserving filter. Then a combination of the noisy unfiltered samples and the less noisy (but biased) filtered samples is added to the low-variance samples in order to form the final image. A novel perpixel blending operator combines both contributions in a way that respects a user-defined threshold on perceived noise. For progressive rendering, this method is superior to similar approaches in several aspects. First, the bias due to filtering vanishes in the limit, making the method consistent. Second, the user can interactively balance noise versus bias while the image is rendering, leaving the possibility to hide filtering artifacts under a low level of dithering noise. Third, the filtering step is more robust in the presence of reflecting/ refracting surfaces and high-frequency textures, making the method more broadly applicable than similar approaches for interactive rendering. The dissertation also contains some optimizations that improve runtime, recover antialiased edges, reduce blurring, and withhold spike noise from the preview images.
The second contribution is the radiance filtering algorithm, another noise reduction method. Again, the basic idea is to exploit spatial coherence in the image and reuse information from neighboring pixels. However, in contrast to image filtering techniques, radiance filtering does not simply filter pixel values. Instead, it only reuses the incident illumination of neighboring pixels in a filtering step with shrinking kernels. This approach significantly reduces the variance in radiance estimates without blurring details in geometry or texture. Radiance filtering is consistent and orthogonal to many common optimizations such as importance, adaptive, and stratified sampling. In addition to the practical evaluation, the dissertation contains a theoretical analysis with convergence rates for bias and variance. It also contains some optimizations that improve the performance of radiance filtering on reflecting/ refracting surfaces and highly glossy surfaces.
The last contribution of this dissertation is a system architecture for exchangeable rendering back-ends under a common application layer in distributed rendering systems. The primary goal was to find a practical and non-intrusive way to use potentially very different rendering back-ends without impairing their strengths and without burdening the back-ends or the application with details of the cluster environment. The approach is based on a mediator layer that can be plugged into the OpenSG infrastructure. This design allows the mediator to elegantly use OpenSG's multithreading and clustering capabilities. The mediator can also sync incremental changes very efficiently. The approach is evaluated with two case studies, including an interactive ray tracer.