Radiance filtering for interactive path tracing

Abstract

We present a method for noise reduction that is especially tailored to interactive progressive path tracing (PT). The idea is to exploit spatial coherence in the image and reuse information from neighboring pixels. However, in contrast to image filtering techniques (e.g. [Schwenk et al. 2012]), we do not simply filter pixel values or samples of outgoing radiance. Instead, we only reuse the incident indirect radiance of neighboring pixels in a radiance estimation step with a shrinking kernel similar to stochastic progressive photon mapping (SPPM) [Hachisuka and Jensen 2009]. This novel approach significantly reduces the variance in indirect lighting without blurring details in geometry or texture. In equal time comparisons we often achieve higher quality than previous approaches. The primary use case of our algorithm is to provide fast, reliable previews of global illumination. It is also consistent, retains the conceptual simplicity of PT, is orthogonal to importance and stratified sampling, and is easy to integrate into existing renderers.