Abstract
This paper introduces the idea of flow field similes for transferring better camera movements and shooting styles from a reference flow field to carelessly shot videos. Given an input video and a reference flow field, we aim to compute a series of homography transformations to warp the input video, so that the flow of the output video will closely resemble the reference flow. The reference flow field may be derived from a real video or synthetically generated. We formulate a nonlinear optimization problem over sparse feature correspondences to find the required transformation for each video frame, through minimizing the difference between the intended reference flow field and the flow field of the output video frame. We show that, by enforcing the flow field of output video to resemble the reference flow field, we are able to create different types of camera movements and shooting styles, from the simplest effect of video stabilization to more complex ones like smooth zooming, anti-blur fast panning, zooming while rotating, tracking shot, and dolly zoom.