Linear and non-linear model for statistical localization of landmarks

This paper presents and compares 3 methods for the statistical localization of partially occulted landmarks. In many real applications, some information is visible in images and some parts are missing or occulted. These parts are estimated by 3 statistical approaches: a rigid registration, a linear method derived from PCA, which represents spatial relationships, and a nonlinear model based upon kernel PCA. Applied to the cephalometric problem, the best method exhibits a mean error of 3.3 mm, which is about 3 times the intra-expert variability.