A Bayesian method for automatic landmark detection in segmented images

Abstract

The identification of landmark points of a figure in an image plays an important role in many statistical shape analysis techniques. In certain contexts, manual landmark detection is an impractical task and an automated procedure has to be employed instead. Standard corner detectors can be used for this purpose, but this approach is not always suitable, as the set of landmark points best representing the figure is not necessarily limited to corners. We present a Bayesian approach for automatic landmark detection, where a set of N landmark vertices is fitted to the edge of a segmented region of an image. We propose a likelihood function for the observed segmented region given the vertices and then use a Metropolis sampler to sample landmark vertices given the observed region. Careful consideration has to be given to the selection of a prior for the distribution of the landmarks.