Bidirectional Clustering for Scalable VPL-based Global Illumination

Abstract

Virtual Point Lights (VPL) methods approximate global illumination (GI) in a scene by using a large number of virtual lights modeling the reflected radiance of a surface. These methods are efficient, and allow computing noise-free images significantly faster that other methods. However, they scale linearly with the number of virtual lights and with the number of pixels to be rendered. Previous approaches improve the scalability of the method by hierarchically evaluating the virtual lights, allowing sublinear performance with respect the lights being evaluated. In this work, we introduce a novel bidirectional clustering approach, by hierarchically evaluating both the virtual lights and the shading points. This allows reusing radiance evaluation between pixels, and obtaining sublinear costs with respect to both lights and camera samples. We demonstrate significantly better performance than state-of-the-art VPL clustering methods with several examples, including high-resolution images, distributed effects, and rendering of light fields.