Towards Fully Dynamic Surface Illumination in Real-Time Rendering using Acceleration Data Structures
Doctoral Dissertation, 2022
Department of Computer Science
Lund University
Abstract
The improvements in GPU hardware, including hardware-accelerated ray tracing, and the push for fully dynamic realistic-looking video games, has been driving more research in the use of ray tracing in real-time applications. The work described in this thesis covers multiple aspects such as optimisations, adapting existing offline methods to real-time constraints, and adding effects which were hard to simulate without the new hardware, all working towards a fully dynamic surface illumination rendering in real-time.
Our first main area of research concerns photon-based techniques, commonly used to render caustics. As many photons can be required for a good coverage of the scene, an efficient approach for detecting which ones contribute to a pixel is essential. We improve that process by adapting and extending an existing acceleration data structure; if performance is paramount, we present an approximation which trades off some quality for a 2–3× improvement in rendering time. The tracing of all the photons, and especially when long paths are needed, had become the highest cost. As most paths do not change from frame to frame, we introduce a validation procedure allowing the reuse of as many as possible, even in the presence of dynamic lights and objects. Previous algorithms for associating pixels and photons do not robustly handle specular materials, so we designed an approach leveraging ray tracing hardware to allow for caustics to be visible in mirrors or behind transparent objects.
Our second research focus switches from a light-based perspective to a camera-based one, to improve the picking of light sources when shading: photon-based techniques are wonderful for caustics, but not as efficient for direct lighting estimations. When a scene has thousands of lights, only a handful can be evaluated at any given pixel due to time constraints. Current selection methods in video games are fast but at the cost of introducing bias. By adapting an acceleration data structure from offline rendering that stochastically chooses a light source based on its importance, we provide unbiased direct lighting evaluation at about 30 fps. To support dynamic scenes, we organise it in a two-level system making it possible to only update the parts containing moving lights, and in a more efficient way.
We worked on top of the new ray tracing hardware to handle lighting situations that previously proved too challenging, and presented optimisations relevant for future algorithms in that space. These contributions will help in reducing some artistic constraints while designing new virtual scenes for real-time applications.
Details
- Thesis Advisor
- Assoc. Prof. Michael Doggett, Lund University
- Thesis Assistant Advisor
- Docent Jacob Munkberg, NVIDIA
- Faculty Opponent
- Assoc. Prof. Veronica Sundstedt, Blekinge Tekniska Högskola
Downloads
- Ph.D. Thesis
- Electronic version
- BibTeX entry
- Coming soon