Bias Correction Applied in Denoising Methods for Monte Carlo Rendering

Abstract

We present a novel approach to correct biases for state-of-the-art denoising methods of Monte Carlo ray tracing. Although current denoising methods have contributed to extremely reduce variances in a rendering output, one of the main problem is that existing methods highly depend on additional feature information e.g. G-buffers, gradients as factors of an edge stop function. If a local image region doesn’t have any data in a set of features, the local will be typically over-smoothed by a denoiser and these amount of smoothness are represented with biases. Especially, measured BRDF e.g. glossy reflection and BSSRDF for subsurface scattering e.g. hair and fur are not defined its features in G-buffers, hence the denoised output image has many biases. In this work, we exactly estimate these biases, which are calculated with images’ gradients and kernel weights of a denoiser according to its definition. Since gradients are noisy in low samples we apply deep autoencoder to estimate reference gradients. Finally, biases are corrected by subtracting them from the denoised output image. Our method improves performance of previous denoising methods both numerically and visually as over-smoothed details are restored.