Date of Award

5-1-2013

Document Type

Thesis (Ph.D.)

Department

Department of Computer Science

First Advisor

Fabio Pellacini

Abstract

As the visual effect and movie industries are striving for realism and high fidelity images, physically based lighting, global illumination, realistic materials, and highly tessellated geometry are gradually accepted and used in movie and game industries. Modern computer graphics has reached an unprecedented level of complexity. As a result, brute-force rendering methods become prohibitively expensive. For decades, computer graphics researchers have been tackling the complexity in rendering by introducing more advanced Monte Carlo integrators, efficient sampling and reconstruction algorithms, effective filtering techniques, sophisticated data structures, and new hardware architectures. In this thesis, we focus on deriving better sampling algorithms to improve the rendering efficiency under complex scene settings. In this context, we explore several areas in computer graphics where sampling algorithms play an important role in improving the overall performance of the renderer: (1). We introduce progressive rendering algorithms as alternatives of full-quality, final rendering under complex settings. We then conduct a user study to evaluate the effectiveness of several progressive rendering algorithms in the context of appearance design tasks. (2). We conduct an investigation into high-quality rendering algorithms for scenes with complex lighting, and propose an efficient many-light algorithm, which renders a few hundred thousand virtual point lights for global illumination, based on matrix slice sampling and light clustering. (3). We investigate importance sampling algorithms for bidirectional scattering distribution functions (bsdfs), and present an importance sampling algorithm for hair bsdf, which can drastically reduce the number of samples required for high quality hair rendering. (4). We look into the problem of out-of-core rendering with massive datasets which cannot fit in the main memory at one time. First, we present an efficient approach to construct out-of-core bounding box hierarchy (BVH). Then, we propose a simple level-of-detail (LOD) model based on point sampling which is inexpensive to compute and compact to store. Finally, we propose a few improvements to the virtual cone tracing algorithm, and present an out-of-core path tracing implementation based on our improved virtual cone tracing algorithm.

Comments

Originally posted in the Dartmouth College Computer Science Technical Report Series, number TR2013-730.

Share

COinS