Fig. 1.

Overview of our keypoint-based fissure segmentation pipeline. We extract a sparse point cloud either in an unsupervised generic (Förstner [16]) or a supervised fissure-specific way (CNN). Then, we apply a geometric segmentation network to assign fissure labels to the points. We reconstruct a dense surface from the segmented points using Poisson surface reconstruction [7] or our point cloud autoencoder