Panorama from Representative Frames of Unconstrained Videos Using DiffeoMeshes

TLDR: The proposed framework for panorama creation consists of four stages: the first stage performs a sparse frame selection based on alignment and blur score, the second stage performs frame alignment using a novel warping model termed as DiffeoMeshes, a demon-based diffeomorphic registration process for mesh deformation, whereas the fourth stage renders the panorama.

Abstract: Panorama creation from unconstrained hand-held videos is a challenging task due to the presence of large parallax, moving objects and motion blur. Alignment of the frames taken from a hand-held video is often very difficult to perform. The method proposed here aims to generate a panorama view of the video shot given as input. The proposed framework for panorama creation consists of four stages: The first stage performs a sparse frame selection based on alignment and blur score. A global order for aligning the selected frames is generated by computing a Minimum Spanning Tree with the most connected frame as the root of the MST. The third stage performs frame alignment using a novel warping model termed as DiffeoMeshes, a demon-based diffeomorphic registration process for mesh deformation, whereas the fourth stage renders the panorama. For evaluating the alignment performance, experiments were first performed on a standard dataset consisting of pairs of images. We have also created and experimented on a dataset of 20 video shots for generating panorama. Our proposed method performs better than the existing state-of-the-art methods in terms of alignment error and panorama rendering quality.