生成当前视点目标图像的快速逆映射算法

A Fast Inverse Warping Algorithm for Generating Novel View

利用极线的属性及深度图像隐含的边界信息 ,提出了一种可以处理非深度连续图像的快速逆映射算法 ,以从多幅参考图像精确合成当前视点目标图像 .该算法分为 3步 :首先由参考图像的深度信息得到图像中的边界 ;然后从参考图像中选出一幅作为主参考图像 ,利用极线的整体匹配特性和对应点在极线上分布的单调性 ,逐个处理目标极线 ,以生成目标图像 ;最后根据其他参考图像填补目标图像中的空洞 .由于在第 2步中只需处理参考图像的边界点对即可得到对应目标极线上所有点的深度信息及其在参考极线上的对应点 ,因此 ,新算法很好地提高了速度 .在填补空洞时 ,利用参考图像的边界信息及所隐含的遮挡关系 。

By taking advantages of the features of epipolar lines and depth discontinuities in reference images, an efficient inverse warping algorithm for generating images of novel views is presented by combining multiple reference images taken from different viewpoints. The method proceeds in three steps. Firstly, the reference images are preprocessed for extracting edge pixels by their displacements. Secondly, a reference image is selected as the primary reference image to be warped inversely epipolar line by epipolar line to render the novel view. Because there exists correspondence between epipolar lines on different images and continuous segments formed by pairs of edge pixels on epipolar lines are order kept, the warping process is very fast by computing only pairs of edge pixels in the reference image to obtain generalized disparity of all pixels in the desired image. Finally, holes in the desired image are filled up through searching the corresponding pixels in other reference images. At this stage, some acceleration techniques are given by utilizing the edge information and occlusion relations between objects in reference images.