PQ kernel: a rank correlation kernel for visual word histograms

by Radu Tudor Ionescu and Marius Popescu

Download the PQK 1.0 software package released under the GNU Public License. The PQK 1.0 software packages contains Matlab and C/C++ implementations of the PQ kernel.

If you use this software (or a modified version of it) in any scientific work, please cite the corresponding paper:

Radu T. Ionescu, Marius Popescu. PQ kernel: a rank correlation kernel for visual word histograms. Pattern Recognition Letters, vol. 55, pp. 51-57, 2015. [BibTeX]

The PQ kernel treats the histograms of visual words as ordinal data, in which data is ordered but cannot be assumed to have equal distance between values. In this case, a histogram will be interpreted as a ranking of visual words according to their frequencies in that histogram.

Usage of the ranks of visual words instead of the actual values of the frequencies may seem as a loss of information, but the process of ranking can actually make PQ more robust, acting as a filter and eliminating the noise contained in the values of the frequencies.

The PQ kernel is roughly 2% better than state of the art kernels, such as the intersection kernel or the Jenson-Shannon kernel, on the Pascal VOC data set.