Fuzzy Shot Clustering to Support Networked Video Databases

Video Databases and video on demand represent an important application of the evolving Global Information Infrastructure. However, video querying involves a lot of user interaction and feedback based query refinement, which can generate large traffic volumes on the network if full video segments are sent. To aid in efficient Video browsing, search and retrieval, we need to find good compact representations for long video sequences. Representative Frames (Rframes) provide such a representation. Extant algorithms use use scene change detection to segment video into shots and pick Rframes. However, scene change detection techniques fail badly in presence of gradual scene changes which are quite prevalent in most videos. We present another way of finding Rframe using fuzzy clustering without dealing with any scene change detection algorithms. Fuzzy clusters provide a more natural approach to this problem since membership of a frame in some particular scene is not binary. This allows us to handle gradual scene changes. We report on our approach, initial experimental resuls, and future plans.