Study On Image-based Rendering

Image-based Rendering (IBR) is a research hotspot both in Computer Graphics and in Computer Vision recently. For very complex scene, we cannot obtain a realistic image by using traditional geometrical modeling techniques of computer graphics. But we can obtain a realistic image by using multiple view geometry theory in computer vision to construct model based on images. This thesis concentrate on the research of the techniques and algorithms of scenes rendering based on photographs, according to the projective geometry framework in Computer Vision.In this thesis, we proposed a novel 3D plenoptic function, which we called border panoramas. The border panoramas are created by constraining viewpoint motion to arbitrary convex region of a plane, and sampling cylindrical panoramas at border of the region point by point. It is different from Concentric Mosaics as follows. The sampling region of border panoramas can be arbitrary convex region, such as rectangle, square, triangle, circle or any other shape. And its sampling manner is sampling cylindrical panorama at border of the region point by point. Border panoramas try to construct a novel flexible representation of 3D plenoptic function by combining the mature panorama technique with the approximate means of ignoring vertical parallax. This method can be used in the VR environment walkthrough system with relatively complex region, and no special sampling device is required.Panoramic mosaics method based on 8-paramter planar homography usually has to overcome the accumulated errors, when a sequence images loops back on itself. The computation load of the common method is very large and theconsistency of homographies cannot be ensured. This thesis presents a simple method, which do not require the consistency of homographies. The method mainly exploits un-calibrated image perspective interpolation technique. So it is of simple calculation and easy realization.Automated panoramic mosaicing method based on cylindrical coordinates is simple and fast, so it is used widely. This thesis presents a robustness-improved fast algorithm for automated panoramic mosaicing. We have mainly solved the following two problems. One is initial parameter estimation, another is small rotation around the center of images. For the former problem, we use resampled small images to compute cross correlation. And we use a 3-parameter model including small rotation to solve the latter problem. This algorithm is appropriate for images taken by hand-held camera, and tripod is no need. Since we use cross correlation match based on small images to estimate the initial value, this algorithm is robust and it also has been used in the ADVR system.We develop a practical system ADVR using the above algorithms. This system can not only support the algorithm based on homography matrix, but also support automated panoramic mosaicing method based on cylindrical coordinates. The browser of panoramas support standard OpenGL interface and can use graphics hardware acceleration function. Comparing to the foreign similar products, ADVR has better image quality and faster computation speed.By taking advantage of parallel lines and orthogonal lines in architecture, the camera internal parameters, rotation and translation can be recovered from a set of un-calibrated images via computing absolute conic and vanishing points. The Euclidean 3D model of architecture (up to a scale factor) can be recovered too. Our method neither requires the camera being calibrated nor need the camera internal parameters being constant.In the realization process of IBR, image inverse mapping and blending techniques are very important. By building a uniform representation on imageVIdisplacement field, we discuss image inverse mapping algorithm in detail. An image iterative blending algorithm is also presented. The results indicate that good effect can be obtained using above algorithms.The research works of IBR have gained many achievements recently. But these achievements still cannot meet the de...