As a computer science PhD candidate and researcher for the Circus Oz Living Archive project, I am investigating how content-based video retrieval systems can enhance search mechanisms within the collection. In this chapter I aim to explore specifically the clues that can assist human annotators in the temporal segmentation of circus performance video into acts. The bulk of the Living Archive collection is comprised of videos of whole performances that are usually over an hour long. To be useful, these videos of whole performances need to be segmented into discrete clips that represent individual acts within the performance; however, this is not a simple task due to the particular characteristics of these videos. Here I identifY the specific clues that indicate an act is likely to have ended, and outline the techniques for the recognition of each clue. The strongest end-of-act clue in the Circus Oz performance video collection is when a clapping sound is detected in combination with one or more other clues: black frames, image changes and audio changes.
↧