Product sub-vector quantization for feature indexing

Abstract

This work addresses the problem of feature indexing to significantly accelerate the matching process which is commonly known as a cumbersome task in many computer vision applications. To this aim, we propose to perform product sub-vector quantization (PSVQ) to create finer representation of underlying data while still maintaining reasonable memory allocation. In addition, the quantized data can be jointly used with a clustering tree to perform approximate nearest search very efficiently. Experimental results demonstrate the superiority of the proposed method for different datasets in comparison with other methods.