Joint-AL: Joint Discriminative and Generative Active Learning for Cross-Domain Semantic Concept Classification

As multimedia data come from a wide variety of domains, each having its distinctive data distributions, cross-domain video semantic concept classification becomes an important task in semantic computing. Its challenge arises from the different distribution (in feature space) of the concept between the source and the target domain, which makes a classifier trained on a source domain perform poorly on a target domain. Active learning can be employed to reuse the existing classifier in order to avoid expensively labeling target domain data for building a new classifier, which queries the labels for a number of most ambiguous samples in target domain and uses these samples to refine the source domain classifier. This discriminative query strategy, used by many traditional active learning methods, could fail if the difference in the feature space distribution of the concept is too large. A generative query strategy is proposed by us in this paper, to deal with large differences between two domains of one semantic concept, which queries samples that are most unlikely to be generated from current distribution. We then present a joint active learning method by adaptively combining the discriminative and the generative query strategies. This method dynamically adapts to the distribution differences and results a hybrid strategy that performs more robustly compared to either single strategy. We evaluate the proposed approaches in cross-domain semantic classification based on TRECVID corpora. The results show the effectiveness of our joint method.