Author:

Keywords:

PSI_VISICS

Abstract:

This thesis deals with the dense multi-view stereo problem. The inherent difficulties which complicate the stereo-correspondence problem are occlusions. Also, we have to consider the possibility that image pixels in different images, which are projections of the same point in the scene, will have different colour values due to non-Lambertian effects or discretisation errors. To tackle these problems we propose a generative model based approach. In this approach, the images are regarded as noisy measurements of an underlying `true' image-function. Also, the image data is considered incomplete, in the sense that we do not know which pixels from a particular image are occluded in the other images. This formulation is equivalent to an inverse inference problem, where the goal is to estimate the factors that have generated the input images. More particular, given a set of images from a scene, we consider the question what would be the most likely image that would have been observed from a particular camera position. To answer this question, we study a global and a local formulation. In a global formulation all possible geometric realisations of a scene are considered and evaluated to find the most plausible realisation. The local formulation takes an initial geometric realisation and refines it in a gradient decent manner. Both formulations are intensely evaluated and their advantages and disadvantages are discussed. Finally, our proposed multi-view stereo algorithm combines both formulations and its performance is illustrated on several real-world examples. We show how the algorithm can generate realistic view interpolations from a virtual camera viewpoint.