Advanced methods for relightable scene representations in image space

The realistic reproduction of visual appearance of real-world objects requires accurate computer graphics models that describe the optical interaction of a scene with its surroundings. Data-driven approaches that model the scene globally as a reflectance field function in eight parameters deliver high quality and work for most material combinations, but are costly to acquire and store. Image-space relighting, which constrains the application to create photos with a virtual, fix camera in freely chosen illumination, requires only a 4D data structure to provide full fidelity. This thesis contributes to image-space relighting on four accounts: (1) We investigate the acquisition of 4D reflectance fields in the context of sampling and propose a practical setup for pre-filtering of reflectance data during recording, and apply it in an adaptive sampling scheme. (2) We introduce a feature-driven image synthesis algorithm for the interpolation of coarsely sampled reflectance data in software to achieve highly realistic images. (3) We propose an implicit reflectance data representation, which uses a Bayesian approach to relight complex scenes from the example of much simpler reference objects. (4) Finally, we construct novel, passive devices out of optical components that render reflectance field data in real-time, shaping the incident illumination into the desired image.