Correlated Topic Models for Image Retrieval

In our previous work [4] we have shown that the representation of images by the Latent Dirichlet Allocation (LDA) model combined with an appropriate similarity measure is suitable for performing large-scale image retrieval in a realworld database. The LDA model, however, relies on the assumption that all topics are independent of each other – something that is obviously not true in most cases. In this work we study a recently proposed model, the Correlated Topic Model (CTM) [1], in the context of large-scale image retrieval. This approach is able to explicitly model such correlations of topics. We experimentally evaluate the proposed retrieval approach on a real-world large-scale database consisting of more than 246,000 images and compare the performance to related approaches.