Bayesian learning for weakly supervised object classification

We explore the extent to which we can exploit interest point detectors for representing and recognising classes of objects. Detectors propose sparse sets of candidate regions based on local salience and stability criteria. However, local selection does not take into account discrimination reliability across instances in the same object class, so we realise selection by learning from weakly supervised data in the form of images paired with their captions. Through experiments on a wide variety of object classes and detectors, we show that modeling object recognition as a constrained data association problem and learning the Bayesian way by integrating over multiple hypotheses leads to sparse classifiers that outperform contemporary methods. Moreover, our learned representations based on local features leave little room for improvement on standard image databases, so we propose new data sets to corroborate models for general object recognition.