Adaptively clustered reflective shadow maps

In this thesis, we present an approach to compute approximate but plausible indirect lighting, a subtopic of global illumination (GI) in computer graphics, in real-time by utilizing the capabilities of modern computer hardware. For years, this topic could only be handled in so-called offline rendering processes that mostly utilize ray-tracing techniques where single images can take up to multiple hours to generate. These offline methods are still largely used in the computer animated movies industry, and only in recent years have ray-tracing methods begun their foray into interactive and real-time computer graphics. However, in contrast to offline methods, real-time computer graphics have traditionally been, and still are (mostly), computed using so-called rasterization methods, by using standardized hardware (graphics cards) and software interfaces. Through the advent of programmable graphics hardware in the early 2000s and seemingly ever-increasing compute capabilities, the topic of global illumination has in recent years finally stepped into the realm of real-time computer graphics. Over the past years, a very active research community produced multitudes of techniques to approximate GI using rasterization-friendly methods. We introduce another variant that is an evolution to a subset of these methods that can be summarized under the term instant radiosity, where indirect lighting is simulated by placing many virtual light sources at the intersection points of light rays with reflecting geometry. Our method, called Adaptively Clustered Reflective Shadow Maps, uses a novel, adaptive clustering approach inspired by k-Means clustering, to reduce the number of required virtual lights without negatively impacting image quality through application of a sophisticated shadowing technique. The aforementioned clustering naturally facilitates temporal coherency by re-using clustering information from the previous frame while simultaneously evaluating its results to counter potential uneven cluster distributions. The results show that our new method exhibits advantages, both performance-wise and image-quality-wise, over previously employed methods.