同心拼图中深度的计算

提出一种新的基于图像绘制的方法：多同心拼图法。该方法将摄像机限制在同一平面绕不同中心的同心圆上旋转,产生多个同心拼图。新的图像由已获得的图像插值产生。多同心拼图法是对Shum提出的单同心拼图法的一种有效改进,其主要特点是能消除单同心拼图法中存在的深度畸变现象。多同心拼图法不需要三维重建过程,具有实现方便、计算简单、能实时产生具有真实感的场景图的优点,在虚拟现实和动画生成方面具有比较重要的应用价值。

Depth Correction in Concentric Mosaics

Image based rendering (IBR) is a novel field which concerns with rendering new images directly from old ones. IBR has many advantages over traditional computer graphics by transcending the complex procedure of modeling. Recently a variety of techniques have been developed in this new field, Shum's Concentric Mosaics(CM) is one of the most promising ones.One of the shortcomings in the standard CM method is its vertical distortion in the rendered new images due to the lack of scene depth information. Although some attempts have been made afterward to handle the problem, which usually involve complex scene modeling, a practical method is still waiting. This paper proposes a novel image based rendering method, namely Multiple Concentric Mosaics (MCM), which can effectively eliminate the vertical distortion problem without scene modeling. Multiple concentric mosaics are created by constraining the camera to rotate on a same plane about different centers, where each rotation generates a single Concentric Mosaics. New images are rendered using the rays captured in MCM. More specifically, at each new view point, the ray of the new view direction will intersect with two capturing circles since we suppose sufficiently many concentric mosaics are captured in MCM. That is to say, for a given new view direction, we can always find a pair of images which are captured in the same view direction. Then from each such pair of images, assuming that the camera's focal length keeps unchanged, the corresponding scene depths can be obtained by some simple algebraic calculations via point correspondences. After obtaining the depth information, photorealistic new images without vertical distortion can be rendered efficiently. The key problem in MCM is to establish point correspondences. The correspondence problem by itself is a difficult one. However, fortunately enough in MCM, this problem can be substantially simplified due to the peculiar epipolar geometry where the two images for point correspondences are always captured in the same view direction. In addition, we also discuss some practical ways for depth correction where only the correspondences of key points, rather than dense points, are carried out. In summary, compared with the standard CM method, the main point of our method is that it can easily get the scene depth information and effectively eliminate the vertical distortion problem. Our method does not require recovering geometric and photometric scene models, and can obtain more photorealistic new images with high computational efficiency. Moreover, like CM, MCM does not need expensive dedicated hardware for image acquisition, and the storage requirement is not too demanding. Finally, experiments show that our MCM method performs satisfactorily.