Evaluating multi-class multiple-instance learning for image categorization

Automatic image categorization is a challenging computer vision problem, to which Multiple-instance Learning (MIL) has emerged as a promising approach. Typical current MIL schemes rely on binary one-versus-all classification, even for inherently multi-class problems. There are a few drawbacks with binary MIL when applied to a multi-class classification problem. This paper describes Multi-class Multiple-Instance Learning (McMIL) to image categorization that bypasses the necessity of constructing a series of binary classifiers. We analyze McMIL in depth to show why it is advantageous over binary MIL when strong target concept overlaps exist among the classes. We systematically valuate McMIL using two challenging image databases, and compare it with state-of-the-art binary MIL approaches. The McMIL achieves competitive classification accuracy, robustness to labeling noise, and effectiveness in capturing the target concepts using smaller amount of training data. We show that the learned target concepts from McMIL conform to human interpretation of the images.