Probabilistic indexing of media sequences

Accurate and fast nearest neighbor search is often required in applications involving media sequences, such as duplicate detection in video collections, music retrieval in digital libraries, and event discovery in streaming documents. Among various related techniques, developing indexing scheme is probably most challenging because of its complexity. This paper documents a novel scheme called HMMH (Hidden Markov Model based Hashing) to facilitate scalable and efficient media sequence retrieval based on advanced hashing algorithm. Main conjecture of our approach is that media sequence's content is complex and the associated dynamic characteristics cannot be ignored. As such, we propose to use hidden Markov model (HMM) for comprehensive media sequence modeling and calculate HMM supervector to represent segments of media sequence. With the novel scheme, more discriminative information about temporal structure can be captured. In addition, the difference of two media sequences is approximated by the Euclidean distance between the associated HMM supervectors. The statistical property enables the proposed HMMH to enjoy good system flexibility - various hashing algorithms (e.g., LSH and SPH) can be applied on HMM supervectors for effective binary code calculation. Our experimental results using both large scale video and music collections demonstrate that the proposed scheme various kinds of advantages over existing techniques.