Reconstruction Techniques for Independent and Correlated Pixel Estimates

Abstract

Monte Carlo (MC) rendering is a widely utilized technique for creating photorealistic images by simulating light interactions in a scene. Due to its randomness from MC integration, MC rendering typically produces pixel estimates that suffer from random noise. Numerous approaches have been proposed to mitigate this noise issue by introducing inter-pixel correlation, such as denoising techniques. While these approaches produce rendered images with less noise, they exhibit different kinds of errors depending on the method. In this dissertation, we propose reconstruction approaches for independent and correlated pixel estimates, aimed at reducing existing errors present in the inputs and thus generating a much-improved image. We first introduce a novel combination function that considers the statistical properties of input images and present a learning-based framework built upon the combination function. Additionally, we propose a self-supervised learning framework for supervised MC denoising, which does not require any pretraining with an external dataset. Finally, we introduce a mathematical denoising framework for unbiased independent and correlated estimates, aimed at reducing variance while limiting denoising bias. We have demonstrated that our proposed methods mitigate different types of remaining errors present in correlated pixel estimates, ultimately improving overall image quality.