{"references": ["L. Wiskott, J.-M. Fellous, N. Kruger, and C. von der Malsburg, \"Face\nrecognition by elastic bunch graph matching,\" IEEE TPAMI, vol. 19, pp.\n775-779, 1997.", "T. Cootes, G. Edwards, and C. Taylor, \"Active appearance models,\" IEEE\nTPAMI, vol. 23, no. 6, pp. 681-685, 2001.", "T. Tamminen and J. Lampinen, \"Sequential Monte Carlo for Bayesian\nmatching of objects with occlusions,\" IEEE TPAMI, vol. 28, pp. 930-\n941, 2006.", "J. Kamarainen, M. Hamouz, J. Kittler, P. Paalanen, J. Ilonen, and\nA. Drobchenko, \"Object localisation using generative probability model\nfor spatial constellation and local image features,\" in Proc. ICCV, 2007,\npp. 1-8.", "C. Doucet, J. de Freitas, and N. Gordon, Sequential Monte Carlo\nMethods in Practice. Springer-Verlag, New York, 2001.", "M. Weber, M. Welling, and P. Perona, \"Unsupervised learning of models\nfor recognition,\" in Proc. ECCV, 2000, pp. 18-32.", "R. Fergus, P. Perona, and A. Zisserman, \"Object class recognition by\nunsupervised scale-invariant learning,\" in Proc. CVPR, 2003, pp. 264-\n271.", "L. Fei-Fei, R. Fergus, and P. Perona, \"A Bayesian approach to unsupervised\none-shot learning of object categories,\" in Proc. ICCV, 2003, pp.\n1134-1141.", "K. Mikolajczyk, B. Leibe, and B. Schiele, \"Multiple object class\ndetection with a generative model,\" in Proc. CVPR, 2006, pp. 26-36.\n[10] S. Lazebnik, C. Schmid, and J. Ponce, \"A discriminative framework for\ntexture and object recognition using local image features,\" Lecture notes\nin computer science, vol. 4170, p. 423, 2006.\n[11] R. Fergus, P. Perona, and A. Zisserman, \"A sparse object category model\nfor efficient learning and complete recognition,\" in Toward Category-\nLevel Object Recognition, ser. LNCS. Springer, 2007, vol. 4170, pp.\n443-461.\n[12] B. Leibe, A. Leonardis, and B. Schiele, \"Combined object categorization\nand segmentation with an implicit shape model,\" Workshop on Statistical\nLearning in Computer Vision, ECCV, pp. 17-32, 2004.\n[13] E. Borenstein, E. Sharon, and S. Ullman, \"Combining top-down and\nbottom-up segmentation,\" in Proc. CVPR Workshop, 2004, pp. 46-53.\n[14] J. Winn and N. Jojic, \"Locus: Learning object classes with unsupervised\nsegmentation,\" in Proc. ICCV, vol. 1, 2005.\n[15] N. Ahuja and S. Todorovic, \"Learning the taxonomy and models of\ncategories present in arbitrary images,\" in Proc. ICCV, 2007, pp. 1-8.\n[16] L. Fei-Fei, R. Fergus, and P. Perona, \"Learning generative visual models\nfrom few training examples: An incremental Bayesian approach tested\non 101 object categories,\" Computer Vision and Image Understanding,\nvol. 106, no. 1, pp. 59-70, 2007.\n[17] J. Daugman, \"Complete discrete 2-D Gabor transforms by neural\nnetworks for imageanalysis and compression,\" IEEE Transactions on\nAcoustics, Speech, and Signal Processing [see also IEEE Transactions\non Signal Processing], vol. 36, no. 7, pp. 1169-1179, 1988.\n[18] R. Neal, \"Probabilistic inference using Markov chain Monte Carlo\nmethods,\" Department of Computer Science, University of Toronto,\nTech. Rep., 1993.\n[19] M. B. Stegmann, \"Analysis and segmentation of face images using\npoint annotations and linear subspace techniques,\" Informatics and\nMathematical Modelling, Technical University of Denmark, Tech. Rep.,\n2002."]}

This paper presents an online method that learns the corresponding points of an object from un-annotated grayscale images containing instances of the object. In the first image being processed, an ensemble of node points is automatically selected which is matched in the subsequent images. A Bayesian posterior distribution for the locations of the nodes in the images is formed. The likelihood is formed from Gabor responses and the prior assumes the mean shape of the node ensemble to be similar in a translation and scale free space. An association model is applied for separating the object nodes and background nodes. The posterior distribution is sampled with Sequential Monte Carlo method. The matched object nodes are inferred to be the corresponding points of the object instances. The results show that our system matches the object nodes as accurately as other methods that train the model with annotated training images.