This paper introduces the Located Hidden Random Field (LHRF), a conditional model for simultaneous part-based detection and segmentation of objects of a given class. Given a training set of images with segmentation masks for the object of interest, the LHRF automatically learns a set of parts that are both discriminative in terms of appearance and informative about the location of the object. By introducing the global position of the object as a latent variable, the LHRF models the long-range spatial configuration of these parts, as well as their local interactions. Experiments on benchmark datasets show that the use of discriminative parts leads to state-of-the-art detection and segmentation performance, with the additional benefit of obtaining a labeling of the object's component parts.
Ashish Kapoor, John M. Winn