Gradient sample argument weighting for robust image region description

Abstract

The weighting of gradient sample arguments for the creation of descriptors of image regions is studied. The descriptors are interpreted as binned and weighted argument kernel density estimates and thereby their defining attributes are identified as the binning rules and the weighting. The weighting is further studied and four different weighting strategies are analyzed. The naive constant weighting is argued to have a poor robustness to image perturbations. As an answer to this, the customary gradient magnitude weighting is motivated. However, the short-comings of this approach are pointed out and two novel weighting strategies are suggested. The first suggested weighting gives a system parameter determining a distinctiveness to robustness trade-off with the customary magnitude weighting being a special case of it. The second suggested weighting gives a similar robustness as the first one, but at a lower computational cost. Finally, the effects of the different weighting strategies are demonstrated with real imagery data and synthetic perturbations.