A Discriminative Data-Dependent Mixture-Model Approach for Multiple Instance Learning in Image Classification Qifan Wang, Luo Si, and Dan Zhang Department of Computer Science Purdue University West Lafayette, IN, USA, 47907-2107 {wang868,lsi,zhang168}@purdue.edu Abstract. Multiple Instance Learning (MIL) has been widely used in various applications including image classification. However, existing MIL methods do not explicitly address the multi-target problem where the distributions of positive instances are likely to be multi-modal. This strongly limits the performance of multiple instance learning in many real world applications. To address this prob- lem, this paper proposes a novel discriminative data-dependent mixture-model method for multiple instance learning (MM-MIL) approach in image classifica- tion. The new method explicitly handles the multi-target problem by introducing a data-dependent mixture model, which allows positive instances to come from different clusters in a flexible manner. Furthermore, the kernelized representation of the proposed model allows effective and efficient learning in high dimensional feature space. An extensive set of experimental results demonstrate that the pro- posed new MM-MIL approach substantially outperforms several state-of-art MIL algorithms on benchmark datasets. 1 Introduction With the pervasion of digital images, automatic image classification has become in- creasingly important. Multiple-instance learning (MIL) [2] is a useful technique in ma- chine learning that addresses the classification problem of a bag of data instances. In multiple instance learning, each bag is composed of multiple data instances associated with input features. The purpose of MIL is to accurately predict bag level labels based on all the instances in each bag with the assumption that a bag is labeled positive if at least one of its instances is positive, whereas a negative bag only contains negative in- stances. In the case of image classification, each image is treated as a bag and different regions inside the image are viewed as individual data instances [15]. The advantage of MIL ascribes to the fact that in training it only requires the label information of a bag instead of individual instances in the bag. However, due to the la- bel ambiguity in the instances, traditional supervised classification methods may not be directly applied to MIL framework. Existing methods in solving MIL problem fall into two categories. The first category is generative model based algorithms, such as axis parallel hyper-rectangles [2], Diverse Density (DD) [9] and Expectation Maximization A. Fitzgibbon et al. (Eds.): ECCV 2012, Part IV, LNCS 7575, pp. 660–673, 2012. c  Springer-Verlag Berlin Heidelberg 2012