Robust Video Mosaicing through Topology Inference and Local to Global Alignment

The problem of piecing together individual frames in a video sequence to create seamless panoramas (video mosaics) has attracted increasing attention in recent times. One challenge in this domain has been to rapidly and automatically create high quality seamless mosaics using inexpensive cameras and relatively free hand motions. In order to capture a wide angle scene using a video sequence of relatively narrow angle views, the scene needs to be scanned in a 2D pattern. This is like painting a canvas on a 2D manifold with the video frames using multiple connected 1D brush strokes. An important issue that needs to be addressed in this context is that of aligning frames that have been captured using a 2D scanning of the scene rather than a 1D scan as is commonly done in many existing mosaicing systems. In this paper we present an end-to-end solution to the problem of video mosaicing when the transformations between frames may be modeled as parametric. We provide solutions to two key problems: (i) automatic inference of topology of the video frames on a 2D manifold, and (ii) globally consistent estimation of alignment parameters that map each frame to a consistent mosaic coordinate system. Our method iterates among automatic topology determination, local alignment, and globally consistent parameter estimation to produce a coherent mosaic from a video sequence, regardless of the camera's scan path over the scene. While this framework is developed independent of the speci c alignment model, we illustrate the approach by constructing planar and spherical mosaics from real videos.