Localized Content Based Image Retrieval with Self-Taught Multiple Instance Learning

There are many scenarios in which multi-instance learning problems may be difficult to solve because of a lack of correctly labeled examples for algorithm training. Labeled examples may be difficult or expensive to obtain because human effort is often needed to produce labels and because there may be limitations on the ability to collect large samples for training from a homogeneous population. In this paper, we present a technique called self-taught multiple-instance learning (STMIL) that deals with learning from a limited number of ambiguously labeled examples. STMIL uses a sparse representation for examples belonging to different classes in terms of a shared dictionary derived from the unlabeled data. This sparse representation can be optimized under the multiple instance setting to both construct high-level features and unite the data distribution. We present an optimization procedure for STMIL along with experiments on localized content-based image retrieval. Our experimental results suggest that, though it learns from a small number of labeled examples, STMIL is superior to standard algorithms in terms of computational efficiency and is at least competitive in terms of accuracy.