Stabilizing image mosaicing by model selection

Abstract

The computation for image mosaicing using homographies is numerically unstable and causes large image distortions if the matching points are small in number and concentrated in a small region in each image. This instability stems from the fact that actual transformations of images are usually in a small subgroup of the group of homographies. It is shown that such undesirable distortions can be removed by model selection using the geometric AIC without introducing any empirical thresholds. It is shown that the accuracy of image mosaicing can be improved beyond the theoretical bound imposed on statistical optimization. This is made possible by our knowledge about probable subgroups of the group of homographies.We demonstrate the effectiveness of our method by real image examples.