No Cover Image

E-Thesis 101 views 11 downloads

Quality Assessment and Variance Reduction in Monte Carlo Rendering Algorithms / Joss Whittle

Swansea University Author: Whittle, Joss O.

DOI (Published version): 10.23889/SUthesis.40271

Abstract

Over the past few decades much work has been focused on the area of physically based rendering which attempts to produce images that are indistinguishable from natural images such as photographs. Physically based rendering algorithms simulate the complex interactions of light with physically based m...

Full description

Published: 2018
Institution: Swansea University
Degree level: Doctoral
Degree name: Ph.D
Supervisor: Mark W. Jones ; Mora, Benjamin
Sponsors / Funders: EPSRC
Grant number: EP/K502935/1
URI: https://cronfa.swan.ac.uk/Record/cronfa40271
Tags: Add Tag
No Tags, Be the first to tag this record!
Abstract: Over the past few decades much work has been focused on the area of physically based rendering which attempts to produce images that are indistinguishable from natural images such as photographs. Physically based rendering algorithms simulate the complex interactions of light with physically based material, light source, and camera models by structuring it as complex high dimensional integrals [Kaj86] which do not have a closed form solution. Stochastic processes such as Monte Carlo methods can be structured to approximate the expectation of these integrals, producing algorithms which converge to the true rendering solution as the amount of computation is increased in the limit.When a finite amount of computation is used to approximate the rendering solution, images will contain undesirable distortions in the form of noise from under-sampling in image regions with complex light interactions. An important aspect of developing algorithms in this domain is to have a means of accurately comparing and contrasting the relative performance gains between different approaches. Image Quality Assessment (IQA) measures provide a way of condensing the high dimensionality of image data to a single scalar value which can be used as a representative measure of image quality and fidelity. These measures are largely developed in the context of image datasets containing natural images (photographs) coupled with their synthetically distorted versions, and quality assessment scores given by human observers under controlled viewing conditions. Inference using these measures therefore relies on whether the synthetic distortions used to develop the IQA measures are representative of the natural distortions that will be seen in images from domain being assessed.When we consider images generated through stochastic rendering processes, the structure of visible distortions that are present in un-converged images is highly complex and spatially varying based on lighting and scene composition. In this domain the simple synthetic distortions used commonly to train and evaluate IQA measures are not representative of the complex natural distortions from the rendering process. This raises a question of how robust IQA measures are when applied to physically based rendered images.In this thesis we summarize the classical and recent works in the area of physicallybased rendering using stochastic approaches such as Monte Carlo methods. We develop a modern C++ framework wrapping MPI for managing and running code on large scale distributed computing environments. With this framework we use high performance computing to generate a dataset of Monte Carlo images. From this we provide a study on the effectiveness of modern and classical IQA measures and their robustness when evaluating images generated through stochastic rendering processes. Finally, we build on the strengths of these IQA measures and apply modern deep-learning methods to the No Reference IQA problem, where we wish to assess the quality of a rendered image without knowing its true value.
Keywords: Computer Science, Stochastic Approximation, Bayesian Statistics, Monte Carlo Methods, Physically Based Rendering, High Performance Computing, Message Passing Interface, Distributed Computing, Machine Learning, Deep Learning, Neural Networks, Convolutional
College: College of Science