Semi-supervised Learning Towards Computerized Generation of Movie Trailers

Generating movie trailers is recognized as an effective way for promoting films and attracting viewers to increase its publicities. Existing techniques, however, are mainly represented by manual process in generating movie trailers, which is time consuming, expensive, and unreliable for the subjectivity of movie trailers. In this paper, we make use of multi-modality features, and propose a semi-supervised learning approach towards automated generation of movie trailer shots without involving any human efforts or manual process. By using a number of well-known movies and their manually-made trailers as the ground truth, we carry out extensive experiments to demonstrate that our approach remains competitive and even outperforms the manually supervised method under certain assessment criteria. Since there exists no commonly agreed evaluation mechanisms for movie trailer generation at present, we show and analyze our experimental results in a number of different angles to support the feasibility and reliability of our proposed approach, providing a good potential for the film industry.