360 Panorama Video Stabilization

This thesis presents a method for stabilizing panorama videos with 360◦ field of view. There are two types of panorama videos: cylindrical one and spherical one. Video stabilization methods for conventional videos do not work properly on these because of the extremely wide angle of view. We observe that rotation needs to be extremely smooth for 360◦ videos to maintain global motion coherency and avoid wobbling. It is also necessary to remove the camera rotation for the virtual reality application since the displaying view should follow the user’s viewpoint when playing the video. Our method decouples the rotation from other motions and applies different strategies for smoothing them. The method first estimates the rotations between frames. After removing rotations, a mesh-based image warping method is used to handle remaining jitters caused by camera translation and parallax. Furthermore, we propose a new type of mesh warping method using 3D icosphere mesh in the spherical case. The desired rotation can then be determined and restored by either heavily smoothing the original rotations or being assigned by the user interaction. Optionally, a tilt correction method can be applied to remove unwanted camera tilt. Experiments show that our method is effective in making steady 360◦ videos.