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. In this paper, we show that such undesirable distortions can be removed by assuming likely subgroups as models and selecting an appropriate one using the geometric AIC. Since no empirical thresholds are involved, out method is suitable for automating the mosaicing process in the form of a software tool. Numerical scheme for optimization is also presented. We demonstrate the effectiveness of our method by real image examples.