Unsupervised deep homography with multi‐scale global attention

Homography estimation serves an important role in many computer vision tasks. Depending heavily on hand-craft feature quality, traditional methods degenerate sharply scenes with low texture. Existing deep homography can handle the low-texture problem but are not robust for overlap rates and/or illumination changes. This paper proposes a novel unsupervised method that simultaneously such and change. Specifically, powerful module, named global transformer contextual encoder (GTCE) is first designed, together correlation to effectively aggregate information reduce matching ambiguity between maps. Moreover, hybrid photo-perceptual loss proposed. The proposed function considers alignment both pixel level perceptual thus helping this network be more adaptive various scenes, including normal cases change cases. results of extensive experiments synthetic real-world datasets demonstrate superiority over current state-of-the-art solutions especially challenging rates, repetitive patterns