26] Y. Tonormura and S. Abe, \content Oriented Visual Interface Using Video Icons for Visual Database System,"

\Scene retrieval method for video database applications using temporal condition changes," in Proc. 5 Summary and future work In this work, two important problems of video databases were addressed. For each problem, a solution based on information readily available in compressed video was presented. As the performance of the shot detection scheme suggests, it is possible to locate shot boundaries in MPEG video with a balanced precision and recall performance by making use of global and local histogramming of I frames. As indicated earlier, it is possible to perform additional processing once the I frames at shot boundaries have been detected to more accurately localize the shot boundaries. Furthermore, since all the results were obtained by using gray-level information only, it is expected that these results can be improved by considering color information additionally. For shot motion characterization, a scheme based on motion vectors of P and B frames was presented. This scheme rst classiies each frame of a shot into one of the distinct motion classes. The collection of labels thus obtained for a shot is next examined to arrive at its motion characterization. In order to learn diierent patterns of motion vectors belonging to diierent motion classes, a supervised decision tree learning scheme was employed. Although, there is a scope for improvement in terms of classiication accuracy, our results demonstrate that it is possible to perform shot motion characterization based on P and B frame motion vectors without the need for computationally expensive optical ow type methods. Since video is truly a multimedia, we are currently investigating methods to process and incorporate compressed audio information for shot boundary detection as well as for shot characterization. Also, we are exploring the issue of shot characterization in terms of other attributes, such as shot range type and the underlying scene. discussed above, the shot motion classiication makes use of the AMIG measure to partition a shot into sub-shots such that each sub-shot has only one dominant motion label for its frames. The next step in shot motion classiication stage is to count to occurrence of each motion label type in a shot or sub-shot. The label that is present more than 50% of the times in a shot is then selected as the shot motion characterization label. If no label type is found to occur more than 50% of the times in a shot, then the shot is declared to …