Inferring semantics with Object Feedback

We propose Object Feedback, a new Bayesian approach to infer semantics in image retrieval. A user provided image typically consists of multiple real world objects with the object of interest being one of them. From a database of object clusters, our system is able to infer the object of interest to the user by finding the representative object cluster having the highest posterior probability. The posterior probability is expressed in terms of the cluster conditional probability and the prior probability of the cluster. In every feedback iteration, we increase the prior probabilities of relevant clusters while reducing that of others. As the cluster priors change in every feedback iteration, the prior probabilities play a dominant role in the calculation of the posterior probabilities of the clusters. We have incorporated the proposed framework in an image retrieval system. We demonstrate the effectiveness of our approach with experiments using a set of categories from the Corel image database.