Binary Video Classification

Large digital video databases will be a fact of life in the very near future, there is no question about it. Just like for the text databases of today, there will be the need for content-based retrieval of this video. Surely, text-based search on script, transcript, and other available text will be an important aspect of this search. But, as we will argue in this paper, text-based search can only bring you so far in this search. Hence, to allow for content-based video search, one has to either perform meticulous manual annotation of the entire video database or design sophisticated computer vision algorithms that do this annotation automatically. The first alternative is not desirable while the second one does not seem to be possible - at lease, not in the near future. So the solution will probably be somewhere in the middle: manual annotation with intelligent computer assistance. Part of this assistance will be video sequence classification in terms of classes with limited semantic meaning.
In this paper, we propose two statistical schemes that classify video sequences into two exclusive classes based upon the characteristics of the underlying shot sequence. That is, the video shot boundaries are first detected and subsequences of shots are then classified using time-series analysis of sequences of shot properties. The two examples given in this paper are the classifications of program versus commercial and action versus non-action shot subsequences. The techniques are computationally simple and efficient, therefor should be useful in assisting manual annotation on early video query stages to narrow down the search. The algorithms have been tested on a variety of video sequences and experimental results are reported.