Venue: Online Talk
Speaker: Rundi Wu, Turing Class (Class of 2016)
Learning generative models of 3D shapes is a key problem in both computer vision and computer graphics. Lately, there has been a steady stream of works on developing deep neural networks for 3D shape generation using different shape representations, e.g., voxel grids, point clouds, meshes, and most recently, implicit functions. However, most of these works produce unstructured 3D shapes, despite the fact that object perception is generally believed to be a process of structural understanding, i.e., to infer shape parts, their compositions, and inter-part relations. In this paper, we introduce a deep neural network which represents and generates 3D shapes via sequential part assembly. In a way, we regard the assembly sequence as a "sentence" which organizes and describes the parts constituting a 3D shape.