Browse Theses

Real-time realistic image synthesis is crucial to many applications, yet continues to present a great challenge in computer graphics. Achieving photorealistic images requires a rendering system to simulate physically correct illumination transport and natural lighting conditions using techniques such as image-based lighting, which represents captured natural illumination by detailed distant environment maps. Existing realistic rendering methods, such as Monte Carlo ray tracing or photon mapping, can accurately simulate physically based rendering effects incorporating realistic lighting and illumination models, but at a significant computational cost that makes them impractical to apply directly in real-time scenarios.

In this dissertation, we present new interactive rendering techniques for accurate, high-quality image synthesis using wavelet-based precomputation. Our underlying approach is based on wavelet approximation of the illumination computation. We begin with environment map rendering, a classic rendering problem that studies the computation of reflections under large-scale image-based lighting. Our main contribution to this problem is the design of fast algorithms for wavelet rotation---rotating globally defined wavelet basis functions into each surface point's local basis set. We present the first computational solution for rotating wavelet basis functions. We precompute rotation matrices that represent the linear transformations between two different wavelet basis sets. Using this approach, we enable fast environment map rendering with high quality reflection effects and interactive manipulation of the lighting, viewpoint, material, and model.

We then turn to the design of a novel system for interactive photorealistic rendering using wavelet-based precomputed light transport. Our system precomputes illumination data about a static scene, enabling interactive rendering that simultaneously captures illumination effects such as glossy surface reflections, realistic shadows, indirect lighting, and translucency. The major challenge with this and similar techniques is that the computation takes place in a high dimensional (6D) illumination sampling space, which results in huge precomputed illumination datasets. We present novel algorithms to efficiently precompute, approximate, and render such large datasets by using wavelet approximation, matrix factorization, and transport accumulation methods. We make several contributions to precomputed light transport. First, we propose compact factored representation of the precomputed illumination data using matrix factorization. Second, we derive efficient transport accumulation algorithms to simulate global illumination effects such as indirect lighting. Third, we introduce the first interactive relighting system that can accurately simulate translucent rendering including both single and multiple scattering effects under image-based lighting. Finally, we present implementation of fast precomputation and rendering algorithms using programmable graphics hardware, achieving substantial performance speedup.