1: \begin{abstract}

2: We study the problem of image alignment for panoramic

3:     stitching. Unlike most existing approaches that are feature-based, our

4:     algorithm works on pixels directly, and accounts for errors across the

5:     whole images globally. Technically, we formulate the alignment problem

6:     as rank-1 and sparse matrix decomposition over transformed images,

7:     and develop an efficient algorithm for solving this challenging

8:     non-convex optimization problem. The algorithm reduces to solving a sequence

9:     of subproblems, where we analytically establish exact recovery

10:     conditions, convergence and optimality, together with convergence rate

11:     and complexity. We generalize it to simultaneously align multiple

12:     images and recover multiple homographies, extending its application

13:     scope towards vast majority of practical scenarios.  Experimental

14:     results demonstrate that the proposed algorithm is capable of more

15:    accurately aligning the images and generating higher quality stitched

16:     images than state-of-the-art methods.

17: \end{abstract}

18: