3D Landmark Localization in Point Clouds for the Human Ear

3D landmark localization plays an important role in many aspects of 3D data processing, from morphometric analysis to the initialization of mesh registration algorithms. In this work we address the problem of landmark localization in 3D point clouds by extending leading 2D landmark localization algorithms to the 3D domain. By leveraging the PointNet++ architecture, we can construct an architecture that is invariant to the ordering of the data. Input point clouds are segmented into background and landmark regions, and offset vectors are calculated within the landmark regions to refine predicted landmark locations. We demonstrate a high landmark localization accuracy, even as the number of points in the input point cloud decreases. By making use of a 3D morphable model as a novel means of data augmentation, improved landmark localization accuracy and consistency can be obtained. We present our results for landmark localization on the human ear.