Image Stitching Using Multiple Homographies Estimated by Segmented Regions for Different Parallaxes

We propose a novel image stitching method using multiple homographies. The method can stitch images having different parallaxes, such as an image that contains distant buildings and close trees. Such images might not be stitched with fine quality by single global homography. Therefore, we select a background homography by analyzing the inliers of the homographies estimated by RANSAC (random sample consensus) for leave-one-out segmented regions, and the left segmented region of the background homography is designated as an object region. Then, the object homography is estimated by the object region. Using these multiple homographies, image stitching and multi-band blending are performed. With this method, images having different parallaxes are stitched with higher visual quality than with other methods using single homography and multiple homographies.