Correspondence Estimation and Image Interpolation for Photo-Realistic Rendering

Free-viewpoint video is a new form of visual medium that has received considerable attention in the last 10 years. Most systems reconstruct the geometry of the scene, thus restricting themselves to synchronized multi-view footage and Lambertian scenes. In this thesis we follow a different approach and describe contributions to a purely image-based end-to-end system operating on sparse, unsynchronized multi-view footage. In particular, we focus on dense correspondence estimation and synthesis of in-between views. In contrast to previous approaches, our correspondence estimation is specifically tailored to the needs of image interpolation; our multi-image interpolation technique advances the state-of-the-art by disposing the conventional blending step. Both algorithms are put to work in an image-based free-viewpoint video system and we demonstrate their applicability to space-time visual effects production as well as to stereoscopic content creation.