Computer graphics generates realistic scene views from geometric models, while computer vision extracts geometric models from real scenes. Image Based Rendering (IBR) is a novel technique of rendering photorealistic images from old ones. With the development of computer techniques, IBR will have many promising applications in Virtual Reality, therefore how to implement the large-scale walkthrough becomes one of the most important goals of IBR. The work in this thesis is a tentative step to this end, and can be summarized as follows: 1. It is well known that the stereo matching is of a crucial problem in computer vision. Matching as many as possible points and avoiding as many false matches as possible are two key aspects of the stereo matching. A new matching technique, named Control Points Based Semi-Dense Matching, is proposed, which can obtain a fairly good number of corresponding points in both texture-rich areas and texture-poor areas and make these matches distribute as evenly as possible. Therefore, it can not only enhance the robustness of interpolation, but also effectively visualize a cloud of reconstructed 3D points. 2. Joint View Triangulation (JVT) Based Image Interpolation is a multiple view representation to handle the visibility issue by consistent parallel triangulation. It is an efficient method to implement the smooth transition between images. Three new concepts, vertical extension, horizontal extension and controllable triangulation, are proposed in this work to improve the JVT structure, and to increase the flexibility of rendering process. 3. Two important prototype systems are implemented in this paper. 1) Concentric Mosaic Based Roaming System. It can realize the smooth roaming, handling well the occlusion and the changes of disparity and rays. 2) JVT Based Image Interpolation System. It can robustly and efficiently realize the automatic linear interpolation between images, and the newly rendered images are photorealistic.
修改评论