3D structure computation from multiple views

3D reconstruction from images is fundamental and has been extensively studied in computer vision. This thesis investigates this problem using multiple images of a rigid 3D scene taken at different viewpoints. We tackle and contribute to several problems in different stages of 3D reconstruction. In the preliminary stage of camera calibration, we work on the problem of camera motion estimation under circular motion. We have developed an effective algorithm that is suitable for a large image sequence. In the 3D reconstruction stage, a dense representation of the scene is computed from calibrated images. We first propose a three-aspect categorization that differen-tiates previous 3D reconstruction methods based on their various properties, such as scene representation, scene prior, visibility computation, optimization method, etc. We conclude that scene representation is most influential among all these properties. We then study two 3D reconstruction problems based on different scene represen-tations. In the first problem, depth maps used to represent the scene are computed from multi-view images. A novel formulation with asymmetrical occlusion treatment is proposed and two approximate optimization methods are developed. Our approach produces comparable results and is more storage and time efficient compared with pre-vious methods, and therefore suitable for multiple input images. The second problem is image-based hair modeling. It is motivated by needs in computer graphics appli-cations. The problem differs a lot from traditional ones due to the high complexity of hair geometry. We developed a practical multi-view based algorithm. It is highly automatic, convenient to use and produces realistic hair models