We investigate projective estimation under model inadequacies, i.e., when the underpinning assumptions of the projective model are not fully satisfied by the data. We focus on the task of image stitching which is customarily solved by estimating a projective warp — a model that is justified when the scene is planar or when the views differ purely by rotation. Such conditions are easily violated in practice, and this yields stitching results with ghosting artefacts that necessitate the usage of deghosting algorithms. To this end we propose as-projective-as-possible warps, i.e., warps that aim to be globally projective, yet allow local non-projective deviations to account for violations to the assumed imaging conditions. Based on a novel estimation technique called Moving Direct Linear Transformation (Moving DLT), our method seamlessly bridges image regions that are inconsistent with the projective model. The result is highly accurate image stitching, with significantly reduced ghosting effects, thus lowering the dependency on post hoc deghosting.
Pairwise stitching example results (“railtracks” image pair)
List of acronyms and initialisms: Baseline-Global homography (via DLT on inliers), SVA-Smoothly Varying Affine, CPW-Content Preserving Warps, DHW-Dual Homography Warps, Autostitch-Panorama tool from the University of British Columbia, Photosynth-Microsoft's ICE Panorama tool, APAP-Our As Projective As Possible Warps.
In order to cogently differentiate the methods, we avoid sophisticated post-processing techniques like seam cutting and straightening, and simply blend the aligned images by intensity averaging such that any misalignments remain obvious. For Photosynth the final post-processed results are presented since the raw alignment is not given.
Panorama example results (“conssite” dataset)
List of acronyms and initialisms: Autostitch-Panorama tool from the University of British Columbia, APAP-Our As Projective As Possible Warps.
You can find some of the datasets we used in our experiments in the following links. The rest of the datasets were taken from the smoothly varying affine warps and from the double homographic warps webpages and authors.