CONF monay:pascal:2006/IDIAP Constructing visual models with a latent space approach Monay, Florent Quelhas, Pedro Gatica-Perez, Daniel Odobez, Jean-Marc EXTERNAL http://publications.idiap.ch/attachments/papers/2006/monay-pascal-2006.pdf PUBLIC http://publications.idiap.ch/index.php/publications/showcite/monay:rr05-14 Related documents the Springer series of Lecture Notes in Computer Science 2006 IDIAP-RR 05-14 We propose the use of latent space models applied to local invariant features for object classification. We investigate whether using latent space models enables to learn patterns of visual co-occurrence and if the learned visual models improve performance when less labeled data are available. We present and discuss results that support these hypotheses. Probabilistic Latent Semantic Analysis (PLSA) automatically identifies aspects from the data with semantic meaning, producing unsupervised soft clustering. The resulting compact representation retains sufficient discriminative information for accurate object classification, and improves the classification accuracy through the use of unlabeled data when less labeled training data are available. We perform experiments on a 7-class object database containing 1776 images. REPORT monay:rr05-14/IDIAP Constructing visual models with a latent space approach Monay, Florent Quelhas, Pedro Gatica-Perez, Daniel Odobez, Jean-Marc EXTERNAL http://publications.idiap.ch/attachments/reports/2005/monay-idiap-rr-05-14.pdf PUBLIC Idiap-RR-14-2005 2005 IDIAP To appear in Springer series of Lecture Notes in Computer Science We propose the use of latent space models applied to local invariant features for object classification. We investigate whether using latent space models enables to learn patterns of visual co-occurrence and if the learned visual models improve performance when less labeled data are available. We present and discuss results that support these hypotheses. Probabilistic Latent Semantic Analysis (PLSA) automatically identifies aspects from the data with semantic meaning, producing unsupervised soft clustering. The resulting compact representation retains sufficient discriminative information for accurate object classification, and improves the classification accuracy through the use of unlabeled data when less labeled training data are available. We perform experiments on a 7-class object database containing 1776 images.