Iterative Training of Discriminative Models for the Generalized Hough Transform

We present a discriminative approach to the Generalized Hough Transform (GHT) employing a novel fully-automated training procedure for the estimation of discriminative shape models. The technique aims at learning the shape and variability of the target object as well as further confusable structures (anti-shapes), visible in the training images. The integration of the learned target shape and anti-shapes into a single GHT model is implemented straightforwardly by positive and negative weights. These weights are learned by a discriminative training and utilized in the GHT voting procedure. In order to capture the shape and anti-shape information from a set of training images, the model is built from edge structures surrounding the correct and the most confusable locations. In an iterative procedure, the training set is gradually enhanced by images from the development set on which the localization failed. The proposed technique is shown to substantially improve the object localization capabilities on long-leg radiographs.