Incremental Bayesian learning of feature points from natural images

Abstract

Selecting automatically feature points of an object appearing in images is a difficult but vital task for learning the feature point based representation of the object model. In this work we present an incremental Bayesian model that learns the feature points of an object from natural un-annotated images by matching the corresponding points. The training set is recursively expanded and the model parameters updated after matching each image. The set of nodes in the first image is matched in the second image, by sampling the un-normalized posterior distribution with particle filters. For each matched node the model assigns a probability for it to be associated with the object, and having matched few images, the nodes with low association probabilities are replaced with new ones to increase the number of the object nodes. A feature point based representation of the object model is formed from the matched corresponding points. In the tested images, the model matches the corresponding points better than the well-known elastic bunch graph matching batch method and gives promising results in recognizing learned object models in novel images.