We address the problem of finding an initial estimation of the location of landmarks on an image, when the landmarks are difficult to distinguish on the image and when the locations are dependent together from external forces such as growth. Our method solves the problem using an adaptive coordinate space where locations are registered. In this space, variability is greatly reduced. A training set is observed to build automatically a mean and a variability model of the landmarks. This model is used to predict the initial estimation on a new image. This method is applied to the difficult problem of the interpretation of cephalograms, with good results.