Density prediction for importance sampling in realistic image synthesis

This dissertation explores new methods based on Monte Carlo integration for generating realistic images of complex environments. Monte Carlo techniques are general and relatively straightforward to implement, which makes them well suited for the global illumination problem. Monte Carlo techniques converge slowly, particularly in complex environments. This work approaches the convergence problem by using carefully chosen probability density functions to create lower variance estimators, a technique known as importance sampling. Another key feature of this work is to exploit precomputation and alternative data structures in order to efficiently implement importance sampling in complex environments. Two successful algorithms based on these concepts are described and implemented. The first of these algorithms computes the contribution of many light sources to a single point statistically, incorporating a light source visibility preprocess to simplify the lighting computations. Next, a radiosity solution is obtained for a manually simplified environment, and this solution is used as the indirect lighting component. When reflectors in the radiosity solution are sufficiently bright, they are reclassified and inserted into the direct lighting computations. This is analogous to replacing a bright (but complex) window with a television monitor that, from a given viewpoint, looks like the window. This does not change how the room looks, but makes computer modeling easier because it reduces complexity. The second of these algorithms attacks both direct and indirect lighting using a light particle tracing preprocess. A five-dimensional data structure (position and direction) maintains regional light flow information and facilitates probability density design for importance sampling. This technique makes few assumptions about the scene, and thus is slow for simple scenes, but does not break down for complex scenes. The light particle data can also be used directly to produce a biased solution with shorter computation time. Both of these techniques produce effective lighting solutions for scenes that have not been feasible for previous algorithms.