Content-Based Video Retrieval Using Video Ontology
Kimiaki Shirahama, Kazuyuki Otaka, Kuniaki Uehara · International Symposium on Multimedia · 2007
In this paper; we aim to efficiently retrieve various kinds of events (e.g. conversation, battle, run/walk and so on) from a archive. To this end, we construct a video which is a formal and explicit specification of events. Specifically, an event is modeled to have 4 dimensions of contents (i. e. Action, Location, Time and Shooting technique). For retrieving such events, concepts in 4 dimensions need to be automatically detected. So, we conduct video data mining to extract semantic from videos. Here, a pattern is a combination of low-level features (e.g. color; motion and audio) associated with events of a certain kind. Thus, patterns can be used to characterize concepts in 4 dimensions of contents. Furthermore, we refine the ontology by extracting new patterns from subspaces of videos, which cannot be retrieved by previously extracted patterns. Finally, we classify each event into genres which potentially contain this event. It is useful for limiting genres from which events of user's interest are retrieved.