SHAPELEARNER: TOWARDS SHAPE-BASED VISUAL KNOWLEDGE HARVESTING

Abstract. The explosion of images on the Web has led to a number of efforts to organize images semantically and compile collections of visual knowledge. While there has been enormous progress on categorizing entire images or bounding boxes, only few studies have targeted fine-grained image understanding at the level of specific shape contours. For example, given an image of a cat, we would like a system to not merely recognize the existence of a cat, but also to distinguish between the cat’s legs, head, tail, and so on. In this paper, we present ShapeLearner, a system that acquires such visual knowledge about object shapes and their parts. ShapeLearner jointly learns this knowledge from sets of segmented images. The space of label and segmentation hypotheses is pruned and then evaluated using Integer Linear Programming. ShapeLearner places the resulting knowledge in a semantic taxonomy based on WordNet and is able to exploit this hierarchy in order to analyze new kinds of objects that it has not observed before. We conduct experiments using a variety of shape classes from several representative categories and demonstrate the accuracy and robustness of our method.