From small to large baseline multiview stereo : dealing with blur, clutter and occlusions
This thesis addresses the problem of reconstructing the three-dimensional (3D) digital model of a scene from a collection of two-dimensional (2D) images taken from it. To address this fundamental computer vision problem, we propose three algorithms. They are the main contributions of this thesis. First, we solve multiview stereo with the o -axis aperture camera. This system has a very small baseline as images are captured from viewpoints close to each other. The key idea is to change the size or the 3D location of the aperture of the camera so as to extract selected portions of the scene. Our imaging model takes both defocus and stereo information into account and allows to solve shape reconstruction and image restoration in one go. The o -axis aperture camera can be used in a small-scale space where the camera motion is constrained by the surrounding environment, such as in 3D endoscopy. Second, to solve multiview stereo with large baseline, we present a framework that poses the problem of recovering a 3D surface in the scene as a regularized minimal partition problem of a visibility function. The formulation is convex and hence guarantees that the solution converges to the global minimum. Our formulation is robust to view-varying extensive occlusions, clutter and image noise. At any stage during the estimation process the method does not rely on the visual hull, 2D silhouettes, approximate depth maps, or knowing which views are dependent(i.e., overlapping) and which are independent( i.e., non overlapping). Furthermore, the degenerate solution, the null surface, is not included as a global solution in this formulation. One limitation of this algorithm is that its computation complexity grows with the number of views that we combine simultaneously. To address this limitation, we propose a third formulation. In this formulation, the visibility functions are integrated within a narrow band around the estimated surface by setting weights to each point along optical rays. This thesis presents technical descriptions for each algorithm and detailed analyses to show how these algorithms improve existing reconstruction techniques.