3D Point Cloud Generative Adversarial Network Based on Tree Structured Graph Convolutions

摘要

The present invention provides an image input unit for receiving an image including object data, and dividing each node of the object data included in the received image into a plurality of child nodes to form a fully connected layer ( A fully connected layer generating unit that generates a fully connected layer, a tree based on an ancestor node corresponding to each node for the positions of a plurality of child nodes for the created fully connected layer for each node A loop term generator that generates a loop term for each node by rearranging the structure into a 3D point cloud, and generates object data included in the received image based on the generated loop term for each node. It has the effect of expressing the geometrical relationship of an object according to the geometrical properties of the graph, that is, a more accurate expression, including a three-dimensional point cloud generator.