Correspondence with category Latent Dirichlet Allocation for image annotation

We present correspondence with category Latent Dirichlet Allocation (corr-c-LDA), a novel probabilistic topic model for the task of image and video annotation. The heart of our annotation model lies in introducing the class label information and assuming the dependence relationships between class label and image feature, as well as class label and annotation words. Instead of modeling the image and annotation words in the formulation of correspondence LDA, our model models the image with class label and annotation words, and tries to avail category information to promote image annotation. We demonstrate the power of our model on 2 standard datasets: a 1791-image subset of UIUC-dataset and a 2400-image LabelMe dataset. The proposed association model shows improved performance over several existing models as measured by F-measure.