Latent Dirichlet Markov Random Fields for Semi-supervised Image Segmentation and Object Recognition

Topic models such as Latent Dirichlet Allocation (LDA) and probabilistic Latent Semantic Analysis have shown great success in segmenting and recognizing the component objects of images. However, such models frequently ignore the spatial relationships among image regions and hence fail to capture important local correlations. In this paper, we introduce the Latent Dirichlet Markov Random Field (LDMRF), a model which improves the spatial coherence of LDA by introducing a Markov random field prior over hidden object class labels. We evaluate the model on a number of semi-supervised joint segmentation and object recognition tasks and compare its performance with LDA. We further demonstrate the advantages of multi-modal feature selection for LDA and LDMRF. Finally, we propose a new combination of variational inference procedures for approximate inference in the LDMRF model and demonstrate its efficacy.