Multiple-Instance Active Learning for Image Categorization

Both multiple-instance learning and active learning are widely employed in image categorization, but generally they are applied separately. This paper studies the integration of these two methods. Different from typical active learning approaches, the sample selection strategy in multiple-instance active learning needs to handle samples in different granularities, that is, instance/region and bag/image. Three types of sample selection strategies are evaluated: (1) selecting bags only; (2) selecting instances only; and (3) selecting both bags and instances. As there is no existing method for the third case, we propose a set kernel based classifier, based on which, a unified bag and/or instance selection criterion and an integrated learning algorithm are built. The experiments on Corel dataset show that selecting both bags and instances outperforms the other two strategies.