Deep Learning Method for Denoising Monte Carlo Renders for VR Applications

Monte Carlo path tracing is one of the most desirable methods to render an image from three-dimensional data thanks to its innate ability to portray physically realistic and desirable phenomena such as soft shadows, motion blur, and global illumination. Due to the nature of the algorithm, it is extremely computationally expensive to produce a converged image. A commonly researched and proposed solution to the enormous time cost of rendering an image using Monte Carlo path tracing is denoising. This entails quickly rendering a noisy image with a low sample count and using a denoising algorithm to eradicate noise and deliver a clean image that is comparable to the ground truth. Many such algorithms focus on general image filtering techniques, and others lean on the power of deep learning. This thesis explores methods utilizing deep learning to denoise Monte Carlo renders for virtual reality applications. Images for virtual reality, or ‘VR’ are composed of both a ‘left eye’ and ‘right eye’ image, doubling the computation cost of rendering and subsequent denoising. The methods tested in this thesis attempt to utilize stereoscopic image data to enhance denoising results from convolutional neural networks