Object segmentation through multiple instance learning

Abstract

An object of interest (OOI) in an image usually consists of visually coherent regions that, together, encompass the entire OOI. We use Multiple Instance Learning (MIL) to determine which regions in an over-segmented image are part of the OOI. In the learning stage, a set of over-segmented images containing, i.e., positive, and not containing, i.e., negative, an instance of the OOI is used as training data. The resulting learned prototypes represent the visual appearances of OOI regions. In the OOI segmentation stage, the new image is over-segmented and regions that match prototypes are merged. Our MIL method does not require prior knowledge about the number of regions in the OOI. We show that, with the coexistence of multiple prototypes corresponding to the regions of the OOI, the maxima of the formulation are good estimates of such regions. We present initial results over a set of images with a controlled, relatively simple OOI. © 2014 Springer International Publishing.