Unsupervised Bayesian Network Learning for Object Recognition in Image Sequences

We present work on the learning of hierarchical Bayesian networks for image recognition tasks. We employ Bayesian priors to the parameters to avoid over-fitting and variational learning. We further explore the effect of embedding Hidden Markov Models with adjusted priors to perform sequence based grouping, and two different learning strategies, one of which can be seen as a first step towards online-learning. Results on a simple data-set show, that the simplest network and learning strategy work best, but that the penalty for the more complex models is reasonable, encouraging work on more complex problems.