360u000e Video Stabilization

摘要

We present a hybrid 3D-2D algorithm for stabilizing 360u000e videonusing a deformable rotation motion model. Our algorithm uses 3Dnanalysis to estimate the rotation between key frames that are appropriatelynspaced such that the right amount of motion has occurrednto make that operation reliable. For the remaining frames, it usesn2D optimization to maximize the visual smoothness of feature pointntrajectories. A new low-dimensional flexible deformed rotation motionnmodel enables handling small translational jitter, parallax, lensndeformation, and rolling shutter wobble. Our 3D-2D architecturenachieves better robustness, speed, and smoothing ability than eithernpure 2D or 3D methods can provide. Stabilizing a video with ournmethod takes less time than playing it at normal speed. The resultsnare sufficiently smooth to be played back at high speed-up factors;nfor this purpose we present a simple 360u000e hyperlapse algorithm thatnremaps the video frame time stamps to balance the apparent cameranvelocity.