Embedding hierarchical clustering in product quantization for feature indexing

Effective indexing is a crucial need for large scale object matching and retrieval. In this work, a novel indexing scheme is presented, that exploits the advantages of hierarchical clustering and product quantization. First, the high dimensional feature space is decomposed into disjointed sub-spaces and the data belonging to each sub-space is separately represented by a hierarchical clustering tree. Second, each tree quantizes a distinct part of an input vector to the closest centroid of a leaf node and the distances for all the pairs of centroids are pre-computed and stored in a lookup table. Finally, searching for a given query is proceeded in parallel between the trees and is performed efficiently in the quantized space using the pre-computed lookup tables. The proposed method has been validated by a number of experiments, demonstrating significant improvements of search performance in comparison with other methods.