Image Stitching of Scenes with Large Misregistration

This paper presents a technique to automatically stitch images captured from different viewpoints. The main contribution of our method is that it allows for largely mismatched content between the adjacent views. To do this, we detect robust keypoints in each of the input images and match them between the adjacent views. Then the correct correspondences are employed to compute the projective matrix through least-square method. After aligning the inputs on the same plane, we use a weight mask to blend the overlapped views with different priorities. This blending strategy keeps the main information of the scene and avoids the interference of the mismatched inputs. We demonstrate the effectiveness of our method on image sets with large misregistration.