Video Upright Adjustment and Stabilization

Abstract

We propose a novel video upright adjustment method that can reliably correct slanted video contents that are often found in casual videos. Our approach combines deep learning and Bayesian inference to estimate accurate rotation angles from video frames. We train a convolutional neural network to obtain initial estimates of the rotation angles of input video frames. The initial estimates from the network are temporally inconsistent and inaccurate. To resolve this, we use Bayesian inference. We analyze estimation errors of the network, and derive an error model. We then use the error model to formulate video upright adjustment as a maximum a posteriori problem where we estimate consistent rotation angles from the initial estimates, while respecting relative rotations between consecutive frames. Finally, we propose a joint approach to video stabilization and upright adjustment, which minimizes information loss caused by separately handling stabilization and upright adjustment. Experimental results show that our video upright adjustment method can effectively correct slanted video contents, and its combination with video stabilization can achieve visually pleasing results from shaky and slanted videos.