Efficient Many-Light Rendering of Scenes with Participating Media

Jan Novák

Ph.D. dissertation

abstract

Many-light rendering has proven to be an efficient and versatile solution for generating synthetic images of virtual scenes. The concept of representing all lighting in the scene using a set of carefully chosen virtual point lights has, however, its specific pitfalls. In particular, contracting all radiant energy into a finite set of infinitesimal emitters leads to singularities, which make high-quality results hard to obtain. Most existing techniques thus simply accept visual artifacts or some form of a systematic error. We present several approaches based on virtual lights that aim at capturing the light transport without compromising quality, and while preserving the elegance and efficiency of many-light rendering.

We first analyze a crucial component of high-quality rendering with virtual point lights, the so-called bias compensation that counteracts the systematic error. By reformulating the integration scheme, we obtain two numerically efficient techniques; one tailored specifically for interactive, high-quality lighting on surfaces, and one for handling scenes with participating media. We then continue investigating the more general problem of solving the light transport in the presence of participating media and present two lighting primitives for many-light rendering: the virtual ray light and the virtual beam light. Representing scattered light with the first primitive significantly reduces the degree of the singularity in the integrand--the major deficiency of rendering with virtual point lights--thereby minimizing the artifacts and allowing for more efficient unbiased computation. The second primitive then avoids the singularity completely by redistributing the emitted energy over the volume of the beam. We demonstrate that these two lighting primitives enable faster convergence and provide better temporal stability than traditional many-light methods based on virtual point lights.

We also address the main bottleneck of (not only) many-light rendering--the visibility testing--and present a novel accelerating structure for fast, approximate ray tracing. Our rasterized bounding volume hierarchies decouple the accelerator from the input geometry by representing the geometry as a collection of hierarchically organized height fields. We show that in addition to fast ray tracing, the hierarchy has a low memory footprint, provides inherent surface parametrization, and natively supports level-of-detail rendering. We demonstrate its advantages in various applications.

downloads

bibtex

@PhDThesis{novak14thesis,
	title 	= {Efficient Many-Light Rendering of Scenes with Participating Media},
	author 	= {Jan Nov\'{a}k},
	year 	= {2014},
	month 	= {May},
	school 	= {Karlsruhe Institute of Technology"}
}