Recognizing panoramas from casual videos

Feng Liu

Abstract

Mosaics imagery is a popular representation for conveying a wide-field
of view of a scene.  Traditional methods require careful planning to
obtain image sources for creating panorama representation.  In this
paper, we propose an automatic method to synthesize panorama from
casually recorded videos.  After pre-processing videos to estimate the
motion models and moving objects, we find video segments suitable for
creating panoramas.  Two criteria are proposed to examine whether a
video segment can be used as sources for panorama.  First, a video
segment shall contains a wide-field of view of a scene to be eligible as
panorama sources.  Second, a video segment shall b emosaicabel, which
means the inter-frame motion shall observe the underlying conditions for
creating panorama.  Based on the video content, two kind of panorama can
be synthesized: scene panorama and activity synopsis.  The former
conveys a wide-view of a scene while the latter conveys the scene
dynamics.  The experiments on various casually recorded videos
demonstrate the effectiveness of the proposed method.