Image Annotation:Pseudo Supervised Bi-latent Dirichlet Allocation

Abstract

In our new model named Supervised Bi-Latent Dirichlet Allocation(pSB-LDA), we explore the value of empirical caption topic proportions in annotation. Having this information. we can predict captions for images easily and efficiently. On the other hand, it turns the problem to supervised topic model that well solved so far. In pSB-LDA, the two data modalities for captions and images are fully trained independently and sequentially, thus caption topics are not effected by training of image model. They therefore reflect the exact distribution of captions in the dataset. We use Logistic regression to represent their cross-relation part by applying a lower bound that famous in variational Bayesian Logistic Regression. Finally, Weshow the outstanding annotation result on 2688-image LabelMe dataset by caption perplexity