From Image-based Motion Analysis to Free-Viewpoint Video
The problems of capturing real-world scenes with cameras and automatically analyzing the visible motion have traditionally been in the focus of computer vision research. The photo-realistic rendition of dynamic real-world scenes, on the other hand, is a problem that has been investigated in the field of computer graphics. In this thesis, we demonstrate that the joint solution to all three of these problems enables the creation of powerful new tools that are beneficial for both research disciplines. Analysis and rendition of real-world scenes with human actors are amongst the most challenging problems. In this thesis we present new algorithmic recipes to attack them. The dissertation consists of three parts: In part I, we present novel solutions to two fundamental problems of human motion analysis. Firstly, we demonstrate a novel hybrid approach for markerfree human motion capture from multiple video streams. Thereafter, a new algorithm for automatic non-intrusive estimation of kinematic body models of arbitrary moving subjects from video is detailed. In part II of the thesis, we demonstrate that a marker-free motion capture approach makes possible the model-based reconstruction of free-viewpoint videos of human actors from only a handful of video streams. The estimated 3D videos enable the photo-realistic real-time rendition of a dynamic scene from arbitrary novel viewpoints. Texture information from video is not only applied to generate a realistic surface appearance, but also to improve the precision of the motion estimation scheme. The commitment to a generic body model also allows us to reconstruct a time-varying reflectance description of an actor`s body surface which allows us to realistically render the free-viewpoint videos under arbitrary lighting conditions. A novel method to capture high-speed large scale motion using regular still cameras and the principle of multi-exposure photography is described in part III. The fundamental principles underlying the methods in this thesis are not only applicable to humans but to a much larger class of subjects. It is demonstrated that, in conjunction, our proposed algorithmic recipes serve as building blocks for the next generation of immersive 3D visual media.