Fig. 6
From: A robust approach to 3D neuron shape representation for quantification and classification
We use two example skeleton graphs (blue and orange) to demonstrate how we embed the skeleton graph. Each node of a skeleton graph is encoded into a feature vector by using graph convolution layers. A fixed length graph level feature vector (global representation) is obtained by graph-level pooling operation of each node feature vector. The discriminator takes inputs both global representation and patch representation to decide whether they are from the same skeleton graph. In this toy example, there will be 14 global-patch pairs