Improving "bag-of-keypoints" image categorisation: Generative Models and PDF-Kernels

Abstract

In this paper we propose two distinct enhancements to the basic
\n''bag-of-keypoints" image categorisation scheme proposed in [4]. In this
\napproach images are represented as a variable sized set of local image
\nfeatures (keypoints). Thus, we require machine learning tools which
\ncan operate on sets of vectors. In [4] this is achieved by representing
\nthe set as a histogram over bins found by k-means. We show how this
\napproach can be improved and generalised using Gaussian Mixture Models
\n(GMMs). Alternatively, the set of keypoints can be represented directly
\nas a probability density function, over which a kernel can be de ned. This
\napproach is shown to give state of the art categorisation performance.