Video corpus construction and analysis

This paper proposes video corpus analysis as a new approach to video handling. The purpose of this approach is to discover frequent and characteristic video expressions from a large amount of video data. A video corpus has been built and currently consists of about 180 hours of MPEG-2 encoded video data, automatically extracted characteristics, and manually tagged attributes. These data include shot boundaries, camera operations, transition time and types between shots, text appearance in video, and thumbnail video frame images. Various tools are developed to enter, analyze, and visualize the video data and attributes. This paper mentions early results; analysis of the video corpus using N-gram statistics of the frame images, probabilities of attributes, and distribution of text appearance timing, reveals some interesting video expressions and usages that can be adopted for video handling