Improving Error Estimation and Convergence for Monte Carlo Denoising Algorithms

Computer graphics and rendering has become pervasive in many industries, in part due to its ability to generate realistic images of complex scenes. Physically based rendering is a general framework for describing the relevant physics of light, in a manner that is tractable for simulation. At the core of this framework is a recursive integral equation, its solution is often estimated using Monte Carlo algorithms in combination with ray tracing. However, such estimates are in general noisy and require that many samples be computed to achieve an accurate and precise solution.

To circumvent this high computational requirement, image denoising is often employed as a post-processing step in conjunction with Monte Carlo rendering. By combining independent pixel estimates, denoising can drastically cut down the number of samples needed to achieve noise-free images. In recent years the state-of-the-art of Monte Carlo rendered image denoising has been dominated by deep learning based solutions, owing to their superior quality over prior approaches especially at low sample counts. Despite significantly improving the efficiency of Monte Carlo rendering, the use of image denoising is not without its drawbacks, as it is prone to introducing bias often perceived as an objectionable amount of blurring in the final image.

In this thesis, we will seek to address the drawbacks and improve upon the stateof-the-art of Monte Carlo denoisers, in ways orthogonal to the choice of underlying denoiser. In particular, we will pay close attention to the error of Monte Carlo denoisers, first seeking to overcome the often objectionable bias of denoised images and then addressing its variance to achieve overall faster error convergence by way of image space adaptive sampling. Lastly, we will present a method for estimating the error of denoised images, paving the way for an automatic stopping criterion for Monte Carlo denoised rendering.