| This thesis deals with the problem of extracting geometrical information from images for use in view synthesis and automatic 3D model creation. The information is extracted from the images alone, without the use of external knowledge in order to overcome the need for calibration. We utilize and demonstrate the use of projective geometry, often in conjunction with tensor notations. The combination of these techniques allows us to accurately express the geometrical constraints that arise. This approach simplifies problem formulation greatly while improving the overall quality of the results. This way, strict camera calibration is not needed and sometimes no calibration is involved at all and so these methods are applicable in a wider range of scenarios. The notion of forgoing explicit camera calibration is gaining wider acceptance nowadays and is supported by many other researchers as well.; We present an algorithm for computing a parametric representation of convex and visual hulls of an object from multiple views with large disparities and use them for novel view synthesis. The method is extremely simple and flexible. The algorithm is silhouette based and requires only projective calibration. Images from any number of views can be gathered to construct a useful representation of the scene that enables model reconstruction and synthesis of new views.; Using the new formulations we extend the classical static configuration to include dynamic scenes as well. The scene may contain many objects moving with fixed velocity along straight lines. The usefulness of such setting is that one can capture the dynamics of a scene using a simple hand-held video camera. We recover the motion in the scene, which may even consist only of moving objects, and use this to synthesize new views from arbitrary viewing positions and arbitrary times.; The well known hierarchy of direct methods approach is extended to include alignment of images induced by a Quadratic surface. This lets us align two views of an object that is not necessarily flat (like a face) in two views without the need for point matches.; Compared to the traditional methods, these show that using less knowledge and assumptions can produce more accurate results while reducing the complexity of the computations. |
