12 th International Workshop of the Multimedia Metadata Community Workshop on Semantic Multimedia Database Technologies co - located

In query-by-example approach, a user can only provide a small number of example shots to represent a query. In contrast, depending on camera techniques and setting, relevant shots to the query are characterized by significantly different features. Thus, we develop a video retrieval method which can retrieve a large variety of relevant shots only from a small number of example shots. But, it is difficult to build an accurate retrieval model only from a small number of example shots. Consequently, the retrieval result includes many shots which are clearly irrelevant to the query. So, we construct an ontology as a knowledge base for incorporating object recognition results into our method. Our ontology is used to select concepts related to the query. By referring to recognition results of objects corresponding to selected concepts, we filter out clearly irrelevant shots. In addition, we estimate a parameter of a retrieval model based on the correlation between selected concepts and shots retrieved by the model. Furthermore, to retrieve a variety relevant shots characterized by different features, we use “rough set theory” to extract multiple classification rules for identifying example shots. That is, each rule is specialized to retrieve relevant shots characterized by certain features. Experimental results on TRECVID 2009 video data validate the effectiveness of our method.