‫תמונות‬ ‫של‬ ‫סדרות‬ ‫התאמת‬ Sequence-to-sequence Alignment

This thesis studies the problem of spatio-temporal alignment of video sequences, i.e., establishingcorrespondences in time and in space between two di erent video sequences of the same dynamicscene. It shows that temporal variations between image frames such as moving objects, changesin scene illumination, or camera ego-motion, are powerful cues for alignment. Such temporalvariations cannot be exploited by standard image-to-image alignment techniques, as they are notcaptured by a single image, but only by a sequence of images. We show that by folding these newtemporal cues and known spatial cues into a single alignment framework, situations which areinherently ambiguous for traditional image-to-image alignment methods are often uniquely resolvedby sequence-to-sequence alignment. This gives rise to a wide range of new video applications.These are discussed in this thesis.The thesis investigates the cases where the sequences are recorded by uncalibrated video cameraswith xed internal and relative external parameters. However, the notion of sequence-to-sequencealignment/matching is more general, and is not restricted to those cases alone.