Searching with quantization package

We provide two implementations of the product quantizer search method described in this paper. This technique allows the search in a large vector dataset indexed in a limited amount of memory.

Matlab version

We provide a simple Matlab implementation of the ADC and IVFADC variants. This toy implementation is provided for the sake of exposition. It is sufficient to reproduce most of the results of our PAMI paper with a reasonable efficiency.

Download Matlab package (Last update: February 5th 2011).

Remarks:

• As a pre-requisite, this package requires the Yael library.
• The evaluation datasets used in our paper are given below.
• This Matlab version is significantly slower than our C implementation. Reporting timings with this implementation is not relevant. We have used our C version to generate all the results (timings+search results) presented in our PAMI paper.

Notice:

This package was written by Herve Jegou in July 2009. Last change: February 2011.
Copyright (C) INRIA 2009-2011.
This version is provided for research purposes only, without any support and with no guarantee.

Pre-compiled library version

We have a quite optimized and well integrated C version, which provides further improved efficiency compared the timings reported in our PAMI paper. Please contact us for commercial usage.

Approximate nearest neighbor search: evaluation datasets

These datasets are now available on a dedicated webpage.