Shot boundary detection using spectral clustering

Daily increase in the number of available video material resulted in significant research efforts for development of advanced content management systems. First step towards the semantic based video indexing and retrieval is a detection of elementary video structures. In this paper we present the algorithm for finding shot boundaries by using spectral clustering methods. Assuming that a shot boundary is a global feature of the shot rather then local, this paper introduces the algorithm for scene change detection based on the information from eigenvectors of a similarity matrix. Instead of utilising similarities from consecutive frames, we treat each shot as a cluster of frames. Objective function which is used as the criteria for spectral partitioning sums contributions of every frame to the overall structure of the shot. It is shown in this paper that optimizing this objective function gives proper information about scene change in the video sequence. Experiments showed that obtained scenes can be merged to from clusters with similar content, suitable for video summarisation. Evaluation is done on different datasets, and results are presented and discussed.