A Bayesian Hierarchical Model for Learning Natural Scene Categories

TLDR

The paper proposes a novel approach to learn and recognize natural scene categories. The method represents scenes as collections of local regions (codewords) grouped into themes, learning both theme and codeword distributions unsupervised without expert annotations. The approach achieves satisfactory categorization performance on 13 complex scene categories.

Abstract

We propose a novel approach to learn and recognize natural scene categories. Unlike previous work, it does not require experts to annotate the training set. We represent the image of a scene by a collection of local regions, denoted as codewords obtained by unsupervised learning. Each region is represented as part of a "theme". In previous work, such themes were learnt from hand-annotations of experts, while our method learns the theme distributions as well as the codewords distribution over the themes without supervision. We report satisfactory categorization performances on a large set of 13 categories of complex scenes.