Unsupervised Object Discovery: A Comparison

摘要: The goal of this paper is to evaluate and compare models methods for learning recognize basic entities in images an unsupervised setting. In other words, we want discover the objects present by analyzing unlabeled data searching re-occurring patterns. We experiment with various baseline methods, based on latent variable models, as well spectral clustering methods. results are presented compared both subsets Caltech256 MSRC2, sets that larger more challenging include object classes than what has previously been reported literature. A rigorous framework evaluating discovery proposed.