Abstract
This paper introduces a novel approach to the fine alignment of images in aburst captured by a handheld camera. In contrast to traditional techniques thatestimate two-dimensional transformations between frame pairs or rely ondiscrete correspondences, the proposed algorithm establishes densecorrespondences by optimizing both the camera motion and surface depth andorientation at every pixel. This approach improves alignment, particularly inscenarios with parallax challenges. Extensive experiments with synthetic burstsfeaturing small and even tiny baselines demonstrate that it outperforms thebest optical flow methods available today in this setting, without requiringany training. Beyond enhanced alignment, our method opens avenues for tasksbeyond simple image restoration, such as depth estimation and 3Dreconstruction, as supported by promising preliminary results. This positionsour approach as a versatile tool for various burst image processingapplications.