从多幅参考图像合成目标图像的逆映射算法

提出了一种从多幅参考图像来合成新视点目标图像的逆映射算法.这种方法不但成功地填补了由于投影区域扩张而产生的第1类空洞,而且还成功地填补了由于空间非深度连续物体相互遮挡而产生的第2类空洞,从而方便地实现了虚拟环境中的漫游.基于物体表面深度的连续性,提出了一种位移预测方法.该方法可以从单幅参考图像获得逆映射过程中所需要的目标图像的位移信息,从而大大提高了算法的效率.与通常的正向映射算法相比,该算法克服了多幅参考图像所带来的计算量成倍增长等问题,而且误差较小.

Combining Multiple Reference Images in a Backward Mapping Algorithm for Novel View Generation

Based on the backward mapping technique, an efficient image-based rendering algorithm for novel view generation from multiple reference images in a static scene is presented in this paper. It successfully fills the holes in the derived image both occured when object surfaces are magnified in the novel view and when occluded surfaces become visible. This algorithm is more efficient by the heuristic technique proposed in this paper which gets displacement values used during backward mapping process from single main reference image under the observation that, the shape of space surfaces often changes smoothly. Comparing with the usual forward mapping methods, this algorithm generates derived images with less errors and tackles the problems associated with multiple reference images.