Illumination for clarity

Understanding the content of real world scenes and presenting the information to human observers is a fundamental task in computer graphics and vision. Traditional methods of imaging a scene using standard photographic techniques only acquire a single sample of each surface's reflectance, specifically the sample at the current camera position and with the current lighting. This information is insufficient for some image understanding tasks. In four novel works, different methods of modifying a scene's illumination to learn more about its contents are investigated. An emphasis is placed on techniques that allow for semi-passive analysis of scenes, meaning they do not create artifacts visible to a human observer. Each of the works examines a different facet associated with adding sources of illumination, and presents the supplementary information in a way which is more intuitive to a human observer. In Material Classification Using BRDF Slices, a static scene is imaged under multiple illumination directions to acquire a more dense sampling of the surface reflectance. These samples are then rectified to correct for differences in surface orientation and are used to disambiguate different materials of similar albedo. In Face Relighting for Video Conferencing, multiple infrared (IR) illumination bases of a video conferencing scene are acquired in real time. A re-illuminated IR scene is generated that reduces distracting lighting effects, and the result is then mapped into the visible domain to be presented to the user. In Depth Matting, IR illumination is used to acquire a low resolution depth estimate of the scene. This information is used in conjunction with high resolution color images to improve the computation of alpha mattes for foreground and background segmentation. In A Context-Aware Illumination, a camera and projector are used to simultaneously analyze and illuminate real world objects. A hand-held IR light is introduced as a novel scene interaction technique which can be used to highlight hard-to-see features directly on the objects themselves, allowing for interactive, real-world viewing by multiple users.