Improving Image Categorization by Using Multiple Instance Learning with Spatial Relation

Image categorization is a challenging problem when a label is provided for the entire training image only instead of the object region. To eliminate labeling ambiguity, image categorization and object localization should be performed simultaneously. Discriminative Multiple Instance Learning (MIL) can be used for this task by regarding each image as a bag and sub-windows in the image as instances. Learning a discriminative MI classifier requires an iterative solution. In each round, positive sub-windows for the next round should be selected. With standard approaches, selecting only one positive sub-window per positive bag may limit the search space for global optimum; meanwhile, selecting all temporal positive sub-windows may add noise into learning. We select a subset of sub-windows per positive bag to avoid those limitations. Spatial relations between sub-windows are used as clues for selection. Experimental results demonstrate that our approach outperforms previous discriminative MIL approaches and standard categorization approaches.