Stochastic Light Cut Exploration

Abstract

Global Illumination provides natural and more realistic illumination of a scene by simulating the bouncing of light between surfaces. Due to its complex nature, it still remains a challenge to simulate this effect with real-time frame rates even on modern-day hardware. Existing methods are noisy and either require extensive filtering or long frame times to converge. In order to increase convergence, Instant Radiosity stores scene illumination from light sources in Virtual Point Lights (VPL), where reflected light rays are considered as new point light sources. Evaluating each VPL for every pixel is slow however. Light Cuts solves this problem by creating a light tree, where clusters of VPLs can be evaluated at once using a representative point light. A light cut is a cut in this tree which determines which clusters are evaluated.

We propose a pipeline which allows for efficient generation of VPLs, light tree generation and light tree sampling for real-time one-bounce global illumination in fully dynamic scenes. VPLs are generated from Bidirectional Reflective Shadow Maps, and we provide an efficient method of building a light tree from these VPLs. Computing full light cuts in real-time is still too costly, thus our work utilizes shallow light cuts which are easier to compute, together with stochastic light tree traversal to reach nodes deeper in the light tree for efficient sampling VPL clusters without sampling correlation. To further reduce computation time, light cuts are computed only for k*k pixel blocks and are blended using a novel stochastic light cut blending strategy.

All these steps make for a global illumination pipeline which is faster and outputs a higher quality image when compared to similar methods.