Support Vector Machines Content-Based Video Retrieval based solely on Motion Information

A new content-based video shot classification method for the purpose of retrieval is proposed, based on the Perceived Motion Energy Spectrum (PMES) descriptor and Support Vector Machines. Using only motion features, we demonstrate the method´s success in learning to separate team sports video shots from all the other videos using real-world material from a TV channel´s archive. We show both the PMES descriptor´s ability to characterize a video shot, and the clear potential of training an SVM to classify any given video into a category, thus moving one more step towards automatic labeling of video.