CREATING AN EFFICIENT QUERY FROM MULTIPLE MULTIMEDIA EXAMPLES Blind for Review

Using multiple examples has become a popular query scenario in multimedia retrieval. This paper explores a unified representation which accumulates various features from different examples to denote a query. Continuous low-level features are quantised into a set of discrete variants. These variants follow a similar distribution as text terms do in a given document collection. Three criteria are compared to justify this projection, including minimised χ, maximised entropy and minimised AC/DC. Statistics similar to text term frequency are computed from these variants for document similarity ranking. Two ranking functions, KL divergence and BM25, are used for multimedia retrieval. The evaluation collection consists of the Corel image set and TRECVid 2006 collection with four low-level visual features. Experimental results show that the overall query performance based on this representation is comparable and in some cases out-performs direct visual feature comparison and the K-median clustering.