Video Stitching Based on Optical Flow
Chunmei Xie, Xiaoyun Zhang, Hua Yang, Li Chen, Zhiyong Gao · 2018
Video stitching remains a challenging problem in computer vision, especially for the widely existed artifacts around moving objects, including parallaxes, ghosts, etc. Traditional methods usually rely on single projective model, which may cause inaccurate correspondences of moving objects in the overlapped region. In this paper, we leverage the optical flow field in the overlapped area, which provides pixel-wise dense projection, so that the artifacts can be drastically reduced. However, since the projection of non-overlapped area is calculated from overlapped area's projection, we propose to automatically select the left or right frame as a reference to avoid the inconsistent transformation when an object moves across the border between overlapped and non-overlapped area. Experimental results demonstrate the advantage of our method over state-of-the-art ones around moving objects.