Motion compensation using spatial transformations with forward mapping

Abstract

Recently, there has been an increasing interest in the use of spatial transformations for motion-compensated prediction. A spatial transformation is a process by which the coordinates of pixels of one image are mapped to new coordinates to form another image. There are two methods by which a spatial transformation can be performed; forward or backward mapping. In most applications, backward mapping is preferred due to its simplicity and fast computation. Forward mapping, on the other hand, is generally avoided since it results in irregularly spaced samples which are difficult to use to reconstruct the desired regularly sampled image. The use of forward mapping spatial transformations, however, has the potential advantage of allowing adaptive motion compensation with no overhead required as compared to backward mapping. In this paper we discuss the use of various techniques for reconstructing regularly sampled images from irregularly spaced samples which result from the use of spatial transformations with forward mapping for motion compensation. A more general problem about the use of spatial transformations for motion compensation is the huge computational load required to find the optimal spatial transformation motion parameters. We also present a new fast search algorithm for refining initial estimates of the spatial transformation motion parameters. We finally evaluate the performance of all the techniques and compare their performance against that of the conventional motion compensation algorithm; the block matching algorithm.