Image Stitching Using Smoothly Planar Homography

It is an important but challenging issue to construct a reasonable seamless image mosaic from images with non-ignorable different viewpoints or multiple distinct planes. The main limitations of existing image-stitching approaches lie in two facts: (a) the multiple plane nature of scenes has not been well considered in the image alignment step, which usually results in obvious misalignments; (b) the ignored alignment errors often lead to broken structures in the seam composition step. To overcome these problems, this paper proposed a smoothly planar homography model for image stitching, by considering the multi-plane geometry of natural scene. First, we integrate local warps estimated in each plane to achieve smoothly plane stitching. Then, we introduce a novel alignment-guided seam composition to handle parallax. Experimental results on a series of challenging data demonstrate that our model achieves the state-of-the-art stitching performance.