Fig. 4
From: A robust approach to 3D neuron shape representation for quantification and classification
Overview of neuron skeleton representation method. Given 3D surface point cloud as input, PointNet++ [23] is used to extract features of the input point cloud. Then a geometric transformation learned by MLP will predict the skeleton points location with their radii. After skeleton points prediction, two simple priors are used to initially connect some skeleton points, and a graph auto-encoder is used to predict all links that connect skeleton points