Experiments with multi-view multi-instance learning for supervised image classification

Abstract

In this paper we empirically investigate the benefits of multi-view multi-instance (MVMI) learning for supervised image classification. In multi-instance learning, examples for learning contain bags of feature vectors and thus data from different views cannot simply be concatenated as in the single-instance case. Hence, multi-view learning, where one classifier is built per view, is particularly attractive when applying multi-instance learning to image classification. We take several diverse image data sets—ranging from person detection to astronomical object classification to species recognition—and derive a set of multiple instance views from each of them. We then show via an extensive set of 10_10 stratified cross-validation experiments that MVMI, based on averaging predicted confidence scores, generally exceeds the performance of traditional single-view multi-instance learning, when using support vector machines and boosting as the underlying learning algorithms.