Estimation of Posterior Marginal Distribution of Each Point in Registration of Point Distribution Model

We propose a non-rigid surface registration method that registers a statistical point distribution model (PDM) to given images and that estimates the posterior marginal distribution of each of the points. We construct a statistical model of the locations of the points based on a set of corresponding points on training surfaces, which is generated by an entropy-based particle system. Given a new image, the method estimates the posterior marginal distributions by means of the non-parametric belief propagation. The proposed method can quantify the accuracy of the each part of registered PDM by referring to the variance of the estimated posterior marginal distribution of each point.