Guided Radiance Transfer in Virtual Light Fields

Realistic image synthesis aims to create images based on synthetic and sampled models that approximate their physical counterparts. Global illumination is a key component that ties together these elements and several techniques have been developed to illuminate virtual environments. Many of these however do not extend to general scenes with arbitrary materials, and yet fewer allow real-time interactivity.

The methods presented as part of my work build on previous work, providing novel methods for storing and computing global illumination in a virtual environment; trading preprocessing and memory usage in favour of interactive rendering. As a first step, the Virtual Light Field extends the notion of light fields for a partial global illumination solution. The scheme parameterises rays into directionally and spatially coherent bundles, using this ray-set for radiance propagation and representation. Though memory and computation requirements are high, the illuminated scene can be viewed interactively.

Building on similar ideas, a significantly different method -- the Hierarchical Virtual Light Field presents an alternative solution to the same problem. We consider global illumination as a sampling problem and develop uniform and adaptive quasi-monte carlo methods for radiance transfer. We present an efficient method for stratified, from-area, non-parametric uniform radiance sampling by efficient solid-angle culling that compares favourably with existing parametric importance sampling methods. This computation also provides a geometric bounds for sampling a transfer of radiance between two surfaces. We employ this sampling scheme to transfer radiance between surfaces represented at various levels of a hierarchy of solid angle discretisations. The method allows for the inclusion of a large set of material types, accounting for all possible light paths for a full global illumination solution. After propagation, the scene can be viewed interactively using the pre-computed data-set, or at higher quality with a more compute-intensive technique.