Path-Space Manifolds for Noise-Free Light Transport

The synthesis of photorealistic images has always been a major challenge in the field of computer graphics and is of great relevance for many applications such as motion picture production, computer games and architectural lighting simulations. Despite our detailed knowledge behind the physical processes of light transport, accurate and efficient simulation is still a challenge and subject to a vast amount of research work. This is due to the fact that light transport is extremely intricate: a single photon, emitted from a particular light source, can interact many times with various materials until it finally reaches the human eye. Current state-of-the-art methods in the field of photorealistic rendering rely on Monte Carlo integration of the incident radiance. As opposed to traditional physical applications, where the solution emerges as the result of one Monte Carlo process, photorealistic rendering is, by principle, different: a separate Monte Carlo process is run for every pixel, resulting in visually unpleasant images. We show that in order to completely eliminate the noise, one either has to resort to rendering an enormous amount of samples per pixel (up to the order of hundreds of thousands) or use noise filtering, where we show that as of now, all methods have their inherent drawbacks. However, our experiments and experience have led us to two key insights that will open up the road towards noise-free photorealistic rendering: Our first key insight is that the most difficult light-transport phenomena usually exist on a low- dimensional manifold that can be specifically sampled exhaustively, resulting in images that are completely noise-free with respect to these difficult light-transport situations. Our second key insight is that noise filtering is a viable way to mitigate noise for simple light transport effects, such as low-frequency illumination on diffuse surfaces. Consequently, the goal of this project is to develop a fundamentally new method that it combines the approaches of (1) exhaustively sampling special path-space manifolds and (2) applying noise filtering to the remaining space. To this end, we will provide a mathematical model to quantify the effectiveness of modern noise-filtering methods. More concretely, the project will make the following contributions to photorealistic rendering: (1) An analysis and visualization of the space of light paths to build a better understanding how most state-of-the-art algorithms can be improved to perform better in challenging light-transport situations. (2) A new class of algorithms that are capable of rendering the most difficult light-transport situations effectively in a fundamentally new way that guarantees that these phenomena always appear completely smooth and converged. (3) A new way of combining noise-filtering techniques with photorealistic rendering algorithms, where each technique is used for the subclass of light paths where it is the most useful.

The generation of photorealistic images is an important aspect in many areas, including movie production and architectural visualisation. In order to mix images of computer- generated objects with real-world film footage, those artificial objects need to look as realistic as possible. Some movies are even completely computer generated, and the imagery needs to look believable in order to immerse the viewer into the story. In the architectural context, it is often important to be able to predict how a room or building will look before it is built. Such pre-visualization images need to resemble reality as closely as possible. All this can be achieved by performing computationally expensive simulations of light transport on the computer. In this project, several approaches have been developed that increase the efficiency of such algorithms and the productivity of people who apply them. Significant advancements have been made in the field of material synthesis and editing. Physically based materials for the generation of photorealistic images are often very complex and have many parameters which make it hard to exactly control the appearance of objects. The developed solutions enable users to quickly synthesise huge amounts of different materials as well as enable adaptation of such materials with common image-based editing operations. This enables fast and efficient generation of materials without the need to run the light simulations for pre-visualization in each iteration of the material design process. Further research in this project was devoted to making the algorithms themselves more efficient. The commonly expensive simulations can be made more efficient and accelerated by carefully guiding it to spend computational resources only on significant parts of the image or scene. We developed methods to use information already available before the simulation is started to efficiently differentiate between more and less important regions. This information is then used during the expensive light simulation to increase its efficiency. These approaches enable to adaptively guide the simulation algorithm to spend computation resources only on significant regions or apply image filtering techniques to increase the image quality after the image was computed.