Abstract

We present MI-CRF, a conditional random field (CRF) model for multiple instance learning (MIL). MI-CRF models bags as nodes in a CRF with instances as their states. It combines discriminative unary instance classifiers and pairwise dissimilarity measures. We show that both forces improve the classification performance. Unlike other approaches, MI-CRF considers all bags jointly during training as well as during testing. This makes it possible to classify test bags in an imputation setup. The parameters of MI-CRF are learned using constraint generation. Furthermore, we show that MI-CRF can incorporate previous MIL algorithms to improve on their results. MICRF obtains competitive results on five standard MIL datasets.