Multiresolution median filtering based video temporal segmentation

In this paper we introduce an enhanced graph partition model, more robust, used for video sequences temporal segmentation. In the first part we examine the formal framework of the shot boundary detection techniques implemented in the recent years, emphasizing the weakness and strength for each method. In the second phase we present our novel algorithm that applies a scale space median filtering on a min-max objective function that optimize the association within each graph cut in order to remove noise caused by camera movement or large object displacement. The algorithm evaluation is done on a subset of movies from the TRECVID 2001 and 2002 video database which demonstrate the improved performance of our method regardless to the noise gender, with precision and recall rates of average 90% and 86 %, respectively. In the end we conclude the paper and present some further implementation.