Abstract:
We present progressive growing of points with tree-structured networks that generates high-fidelity point cloud. Because point cloud data lacks the information of inherent topology or connectivity between neighboring points, the data generated from deep neural networks usually fails to faithfully produce local details. Inspired by the recent success of the progressive generation of images and curriculum learning, we suggest that the hierarchical structure of the tree-based network architecture can endow contextual information to enforce the progressive generation of point clouds. When the tree-structured network is incrementally trained by progressively adding the subsequent layers of depth, the quality of generated point cloud is superior to the data generated by the same network structure with naïve end-to-end training. Furthermore, our pipeline simultaneously learns the hierarchical structure within the data set and finds a consistent spatial decomposition of 3D shapes by coherently positioning the nodes with the same ancestors. Extensive experiments show that our method can produce a high-fidelity shape when applied to shape generation and completion as well as auto-encoding point clouds.
