Stitching Videos From Unstructured Camera Arrays With Rectangular Boundaries

This paper presents a novel warping based method to stitch videos from unstructured camera arrays. Our approach adopts a two-step energy optimization for video stitching. In the first step, we perform an initial stitching on keyframes, and then extract the boundary vertices and warped vertices as constraints for further optimization. In the second step, we design a global optimization to effectively propagate the stitching from the keyframe to other frames while ensuring the feature alignment, boundary regularity and temporal coherence. The optimization can be efficiently solved by a linear system, and the final stitching results are produced by warping and blending. Experimental results and comparisons show that our method can efficiently stitch multiple videos from unstructured camera arrays, and outperforms state-of-the-art methods.