Firefly detection with half buffers

Abstract

Fireflies, or noise spikes, are overly-bright pixels out of place compared to neighboring pixels, which are a common artifact in Monte Carlo ray traced images. They arise from low-probability events, and would be resolved in the limit as more samples are taken. However, these statistical anomalies are often so far out of the expected range that the time for them to converge, even barring numerical instabilities, is prohibitive. Aside from the general problem of fireflies marring a rendered image, their difference in color and variance values can cause problems for denoising solutions. For example, the distance calculation for non-local means filtering [Buades et al. 2005] presented in Rousselle et al. [2012] is not robust under extreme differences in variance. This paper addresses removing these fireflies to improve both the rendered image on its own, and making the available data more uniform for denoising solutions. This paper assumes a denoising framework that makes use of half buffers and pixel variance, such as set forth in Rousselle et al. [2012] and Bitterli et al. [2016]. The variance provides better data than the color channels for determining which pixels do contain fireflies, whereas the half buffers provide some assurance that the detected firefly is not an expected highlight in the rendered image.