Adaptive Sampling for Geometry-aware Reconstruction Filters

We present an adaptive sampling scheme for Monte-Carlo-based renderers with the aim to support geometry-aware filtering techniques for interactive computation of global illumination. While sophisticated filtering techniques for homogeneous areas can already produce high-quality results with as few as one sample per pixel, these approaches lack the ability to filter sufficiently in the vicinity of complex geometric structures. The result are visible artifacts in the final rendering result. Our sampling scheme distributes the samples for the indirect illumination in the image plane according to the necessity of a geometry-aware filtering. We show how to implement our scheme efficiently on current graphics hardware and how to combine it with a sophisticated filtering in order to achieve high-quality interactive frame rates for global illumination simulations. The resulting computational overhead is only in the range of a few milliseconds, making our approach suitable for real-time implementations.