inverted-quantized k-means
C++ Matlab Other
Switch branches/tags
Nothing to show
Clone or download
Latest commit 666889f Mar 1, 2016
Failed to load latest commit information.
anim readme Mar 1, 2016
cluster inputID Sep 14, 2015
config c2_test/eval_iter Sep 18, 2015
disp minor Apr 18, 2015
exp/small xio May 24, 2015
lib drvq stuff Sep 17, 2015
quant c2_test/eval_iter Sep 18, 2015
sub camera Sep 12, 2015
test eval_iter Sep 22, 2015
util experimental Sep 15, 2015
.gitignore purge Apr 14, 2015
LICENSE license Sep 22, 2015 readme Mar 1, 2016


iqm stands for inverted-quantized k-means (IQ-means) and implements a method for fast approximate clustering.


iqm has a 2-clause BSD license. See file LICENSE for the complete license text.

Directory structure

iqm constists primarily of Matlab .m files, but also includes a number of .cpp files which are compiled as mex files to interface Matlab. The directory structure is:

/anim      animation of algorithm iterations on 2D example
/config    project configuration, including datasets, paths etc.
/disp      text/graphical display
/lib       generic utilities library
/quant     full set of learning, encoding, inversion and search methods
/sub       underlying code handling subspaces
/test      main test entry points
/util      specific utilities for this application


iqm requires Matlab, an appropriate C++ compiler to compile mex files (e.g. gcc), yael library, and xio library.


After installing yael, choose a location on disk, say home, where to store datasets and output. Make one subfolder for each dataset with the name of the dataset, say sift, and copy the raw dataset files in a subfolder named raw. For instance, download the SIFT1M dataset and unzip its contents into


iqm uses the specific filenames and file formats as given in the link above. A number of additional folders for output will be created automatically under home/sift/.

Finally, edit /config/local.m and adjust the location of home and yael to your local settings. Now you can try any of the samples under /test/. The mex files should compile automatically.


Please cite the following paper if you use this software.

Y. Avrithis, Y. Kalantidis, E. Anagnostopoulos, and I. Z. Emiris. Web-scale image clustering revisited. In Proceedings of International Conference on Computer Vision (ICCV 2015), Santiago, Chile, December 2015.