Efficient Locality Sensitive Clustering in Multimedia Retrieval

Clustering is fundamental in multimedia retrieval. For example, visual features of high dimensionality are extracted and clustered for image content analysis in image retrieval, scene classification and object retrieval applications. Existing clustering methods suffer from the curse of dimensionality when data are high dimensional and especially in large scale. Locality Sensitive Hashing (LSH) [1] is proposed to remove this problem in high-dimensional indexing, which is the most popular indexing schema in multimedia. We propose an approximate clustering method named Locality Sensitive Clustering (LSC) for high dimensional data in large scale situations. LSC uses pivots to estimate similarities between data points and generates clusters based on the Locality Sensitive Hashing scheme. Experiments on open datasets show that LSC achieve significant improvement on clustering efficiency (i.e. in magnitudes) with little loss of accuracy compared to the state of the art methods.