Codebook+: A New Module for Creating Discriminative Codebooks

Ziming Zhang, Syin Chan, Liang-Tien Chia

Abstract: In this paper, we introduce a new module, Codebook +, into a classical framework which combines bag-of-words image representation with probabilistic latent semantic analysis (pLSA) for unsupervised object categorization. This new module makes the framework less sensitive to the image sampling methods as well as improves its performance. In this module, we create a new codebook based on the discriminability of each codeword in the original codebook for different categories. In our experiments, we compare the classification results of the framework with and without Codebook + using five different image sampling methods.