We present a method for surface recovery in partial surface scans based on a statistical model. The framework is based on multivariate point prediction, where the distribution of the points are learned from an annotated data set. The training set consist of surfaces with dense correspondence that are Procrustes aligned. The average shape and point covariances can be estimated from this set. It is shown how missing data in a new given shape can be predicted using the learned statistics. The method is evaluated on a data set of 29 scans of ear canal impressions. By using a leave-one-out approach we reconstruct every scan and compute the point-wise prediction error. The evaluation is done for every point on the surface and for varying hole sizes. Compared to state-of-the art surface reconstruction algorithm, the presented methods gives very good prediction results.
Lecture Notes in Computer Science: Meshmed 2012 Proceedings, 2012, p. 49-58
Main Research Area:
Lecture Notes in Computer Science
15th International Conference on Medical Image Computing and Computer Assisted InterventionMedical Image Computing and Computer Assisted Intervention, 2012