Learning Semantic Visual Concepts from Video

Increasing amounts of digital video data have become available with the rapid growth in video technology. As a result, there is a great need for automatic extraction of concepts or events of interest from video. In this paper, we present an approach for learning concepts from video. The approach consists of three steps. In the first step, video shot boundaries are detected, and from these shots key frames are extracted, which are representatives of the shots. In the second step, key frames are segmented and a variety of features are computed. In the third step, a classification by feature partitioning method is employed for learning different semantic concepts. The results are presented for successfully learning semantic concepts such as ocean, mountain, people, and building from a variety of digital videos.