Image Inpainting using Orthogonal Viewpoints and Structure Consistency in Manhattan World
Irawati Nurmala Sari, Yuto Urano, Weiwei Du · 2021
This paper proposes a fast and straightforward method for restoring a damaged hole in the Manhattan world, which has three orthogonality with one unknown camera to produce a perspective view that can foreshorten planes of similar structure. There are two essential guidance for image inpainting algorithms: Orthogonal Viewpoints and Structure Consistency. First, due to perspective view, we build orthogonal viewpoints from an image by determining the vanishing points of different planes. Second, we present structure consistency for reconstructing unknown regions of possibly foreshortened planes due to the displacement vector of repeated structures in perspective planes. For image inpainting, PatchMatch [1] algorithm handles matching structures based on two prior-primary guided methods. Our experiment results prove that our approach completes image inpainting in challenging scenes, such as perspective views, repetitive structures, and facades.