360 Video Stabilization

摘要

We present a hybrid 3D-2D algorithm for stabilizing 360 videousing a deformable rotation motion model. Our algorithm uses 3Danalysis to estimate the rotation between key frames that are appropriatelyspaced such that the right amount of motion has occurredto make that operation reliable. For the remaining frames, it uses2D optimization to maximize the visual smoothness of feature pointtrajectories. A new low-dimensional flexible deformed rotation motionmodel enables handling small translational jitter, parallax, lensdeformation, and rolling shutter wobble. Our 3D-2D architectureachieves better robustness, speed, and smoothing ability than eitherpure 2D or 3D methods can provide. Stabilizing a video with ourmethod takes less time than playing it at normal speed. The resultsare sufficiently smooth to be played back at high speed-up factors;for this purpose we present a simple 360 hyperlapse algorithm thatremaps the video frame time stamps to balance the apparent cameravelocity.