Experiments on Selection of Codebooks for Local Image Feature Histograms

Abstract

Histograms of local features have proven to be powerful representations in image classification and object detection. In this paper we experimentally compare techniques for selecting histogram codebooks for the purpose of classifying 5000 images of PASCAL NoE VOC Challenge 2007 collection. We study some well-known unsupervised clustering algorithms in the task of histogram codebook generation when the classification is performed in post-supervised fashion on basis of histograms of interest point SIFT features. We also consider several methods for supervised codebook generation that exploit the knowledge of the image classes to be detected already when selecting the histogram bins.