Fast, Automated 2D Estimation and Error Reduction by Model Selection

In the building of 2D image mosaics using homographies, numerical instability results in errors during registration which results in distorted images. In many occasions, irrespective of the nature of the transformation, a projective transformation is commonly used to compute the homography. In this paper, we demonstrate that by selecting the appropriate transformation to register the images, the error can be reduced even if overlapping area of the images are small. In addition we used the method of projection on convex sets (POCS) to improve the sub-pixel accuracy which reduces the error in the panormic image. It is shown that image mosaicing can be improved by knowledge of the scene which allows selection of the appropriate subgroup of homographies.