Enhacing Image Mosaicing with Adaptive Local Homographies

Image alignment is usually the first move in modern image stitching pipeline; and utilizing a global homography to align two images is justified on the condition that only pure rotation concatenates two views or the whole scene is planar. Ghosting effects frequently happen because aforementioned condition is rarely satisfied in practice; so we propose adaptive local homographies to improve image alignment further based on as-projective-as-possible warp (APAP). Specifically, we introduce an adaptive inlier selection scheme for fitting local homographies while APAP adopts moving DLT to do this with improper global inliers, so our warp can be more ”local” than APAP while still globally projective. As experiments show our warp moderately outperforms APAP in alignment accuracy and deblurring especially in the case of large parallax, which leads to a powerful alternative to moving DLT with acceptable overhead.