Bayesian view synthesis for video stitching

Jianmei Su, Hong Cheng, Lu Yang, Ao Luo

Abstract: In this paper, we propose a Bayesian video stitching approach in multiview camera configuration considering the potential alignment errors with the imprecise optical flow estimation. The emphasis is on the artifacts eliminating around occlusions and motion objects edges. Traditional video stitching approaches ignored alignment errors leading to significant artifacts. In this paper, we first analyze to account for artifacts in the stitched video by imprecise optical flow estimation. Then, we derive and apply the spatial-temporal reliability provided by the error approximation in video stitching. We synthesis the final stitched result by solving a Maximum a Posterior(MAP) problem using graph cuts. The experimental results show the outperformance against the state-of-art video stitching approaches.