Rendering from unstructured collections of images

Abstract

Computer graphics researchers recently have turned to image-based rendering to achieve the goal of photorealistic graphics. Instead of constructing a scene with millions of polygons, the scene is represented by a collection of photographs along with a greatly simplified geometric model. This simple representation allows traditional light transport simulations to be replaced with basic image-processing routines that combine multiple images together to produce never-before-seen images from new vantage points. This thesis presents a new image-based rendering algorithm called unstructured lumigraph rendering (ULR). ULR is an image-based rendering algorithm that is specifically designed to work with unstructured (i.e., irregularly arranged) collections of images. The algorithm is unique in that it is capable of using any amount of geometric or image information that is available about a scene. Specifically, the research in this thesis makes the following contributions: * An enumeration of image-based rendering properties that an ideal algorithm should attempt to satisfy. An algorithm that satisfies these properties should work as well as possible with any configuration of input images or geometric knowledge. * An optimal formulation of the basic image-based rendering problem, the solution to which is designed to satisfy the aforementioned properties. * The unstructured lumigraph rendering algorithm, which is an efficient approximation to the optimal image-based rendering solution. * A non-metric ULR algorithm, which generalizes the basic ULR algorithm to work with uncalibrated images. * A time-dependent ULR algorithm, which generalizes the basic ULR algorithm to work with time-dependent data.