Representing images with χ2 distance based
histograms of SIFT descriptors
Histograms of local descriptors such as SIFT have proven to be powerful representations of image content. Often the histograms are formed using a clustering algorithm that compares the SIFT descriptors with the Euclidean distance. In this paper we experimentally investigate the usefulness of basing the comparisons of the SIFT descriptors on the χ2 distance measure instead. The modified approach results in improved image category detection performance when it is incorporated into a Bag-of-Visual-Words type category detection system.