Unsupervised learning of spatial structures shared among images

Learning from unlabeled images that contain various objects that change in pose, scale, and degree of occlusion is a challenging task in computer vision. Shared structures embody the consistence and coherence of features that repeatedly cooccur at an object class. They can be used as discriminative information to separate the various objects contained in unlabeled images. In this paper, we propose a maximum likelihood algorithm for unsupervised shared structure learning, where shared structures are represented as the strongly connected clusters of consistent pairwise relationships and shared structures of different order are learned through exploring and combining consistent pairwise spatial relationships. Two routines of sampling data, namely densely sampling and sparsely sampling, are also discussed in our work. We test our algorithm on a diverse set of data to verify its merits.