Learning spatial prior with automatically labeled landmarks

We propose a method of automatically labeling landmarks on target images, which are used for training a constellation model to recognize general object class. First, we randomly sample local features (parts) and generate hierarchical representations of images in a similar way to the ¿standard model¿ of visual cortex. Second, we pick out a unique location of each part among those local maxima in S2 layers by a matching procedure. Third, we model the spatial relations among parts as a sparse GMRF (Gaussian Markov random fields) graph, and learn the links by a lasso-based approach. Object localization in new images proceeds by maximizing the posterior of an object observed at a particular configuration. Our model is a thoroughly automatic scheme to perform ¿feature binding¿. Experimental results on the CalTech101 database demonstrate that the proposed algorithm locates the components more precisely and outperforms the ¿standard model¿ in object detection.