Deprecation notice (June 2020)

Please note that I am no longer mantaining this code, and I discourage you from using Matlab.

You can run stacked quantizers and many other quantization methods in Julia using Rayuela.jl

---

Stacked Quantizers

This is the code for the paper

Julieta Martinez, Holger H. Hoos and James J. Little: Stacked Quantizers for Compositional Vector Compression, available at: http://arxiv.org/abs/1411.2173

This code was mostly written by Julieta Martinez.

Datasets

The demo requires you to download the convnet1m-128 dataset, and put it into a /data folder at the top directory.

• The datasets for training, queries and database are available in: features_m_128.mat
• The ground truth is available in: features_m_128_gt.mat

SIFT1M and GIST1M can be downloaded from INRIA.

Demo

For a demonstration of approximate nearest neighbour search

1. Compile search utilities by running compile.m from the top directory.
2. Run demo.m from the top directory as well.

Citation

If you find this code useful, please consider citing our paper:

Julieta Martinez, Holger H. Hoos, and James J. Little.: Stacked Quantizers for Compositional Vector Compression. arXiv preprint arXiv:1411.2173 (2014).

Acknowledgements

• The code under /OPQ was taken from Cartesian k-means by Mohammad Norouzi
• The code that draws the beautiful recall@N plots under /SQ/util/eval_recall_vs_sel.m is by Kaiming He