No.121 May 2010

Research Presentations

  • Fast Method of Shot Boundary Detection Based on Sequential Decision Procedure
    Yoshihiko KAWAI, Hideki SUMIYOSHI and Nobuyuki YAGI
    ↓summary

    summary
    Shot boundary detection is defined as processing of dividing video data into short video segments based on shot boundaries which are points in time when TV cameras are switched. Shot boundary detection is one of the most fundamental processes in video analysis. In this paper, we propose a novel method which enables precise, high-speed detection by omitting the processing of frames that are clearly not shot boundaries, and by analyzing various features sequentially only for the parts of the video that are likely to contain shot boundaries. An evaluation on actual broadcast video resulted in a recall rate of 90.4% and a precision rate of 92.8% on average. About 425 minutes of test data was processed in 208 seconds (excluding the MPEG-1 decoding time), or 1/123rd of the real time.
  • Play Classification of Baseball Broadcast Video Scenes Based on Symbol Sequenced Scene Focusing on Post Pitch Shot
    Takahiro MOCHIZUKI, Mahito FUJII, Nobuyuki YAGI and Kouichi SHINODA
    ↓summary

    summary
    This paper describes a method for automatically classifying baseball video scenes into playclasses(i.e., homerun, single, walk, etc.).Our method is based on a technique to simplify a videointerval using a set of rectangles with image features and motion information. The basic unit for simplification is a shot. For the second shot of each scene that includes significant information for play-classification, a partial shot generated by dividing the shot is used as a processing unit. The scenes used for training are expressed as sequences of symbols based on the simplified data for shots and partial shots.;Play-class-unknown”baseball scenes are assigned one of the playclasses by using discrete hidden Markov models that have been trained with the training symbol sequences for each kind of play-class. An experiment using videos of seven Major League Baseball games produced good results, demonstrating that this method can automatically classify scenes with high accuracy.
  • Automated Production of TV Program Trailer Using an Introductory Text from Electronic Program Guides
    Yoshihiko KAWAI, Hideki SUMIYOSHI and Nobuyuki YAGI
    ↓summary

    summary
    Video abstraction is defined as producing shorter video clips from the original video and it is one of the most efficient methods for retrieval from large video archives. We propose an automated method of producing TV program trailers. Our method employs introductory text from an electronic program guide, which is a short description of the program highlights. We extract closed caption sentences that have the highest similarity for each introductory sentence and then connect the corresponding video segments to make the trailer. A Bayesian belief network is used to calculate the similarity. The proposed method was used to generate trailers for actual TV programs, by which their effectiveness was verified.
  • A Method of Generating of Image based Quizzes from News Video Archives
    Masanori SANO, Nobuyuki YAGI, Norio KATAYAMA and Shin’ichi SATOH
    ↓summary

    summary
    We describe a method of generating quizzes from a news video archive. An image-based multiple-choice-quiz was formulated with three sub-tasks. These tasks include selecting an appropriate image in the quiz, selecting an appropriate sentence describing the image, and generating multiple choice questions on the image. We describe the engineering method for each sub-task and our tests of them. The effectiveness of the method was demonstrated in our experiments. Finally, we discuss what needs to be done to improve the accuracy and quality of generating quizzes.