Extending Deep Interactive Evolution with Graph Kernel for 3D Design

Abstract

DeepIE3D, a recent research, enables users to generate their favorite 3D structures by combining GAN and IEC. However, due to the stochastic nature of IEC, it is very difficult to evolve and generate specific structure, even under human guidance. To solve this problem, the system needs to pick out 3D structures that are desirable to users, and for this purpose, it is necessary to define some kind of similarity measure to extract advantageous features from selected structures. We would like to propose to use DeepIE3D with graph kernels. In this work, we represent planes/chairs as graphs and used Weisfeiler-Lehman graph kernels to implement recommendation system. The result shows that the proposed method is superior in generating specific types of planes/chairs and the proposed similarity calculation method are very intuitive from a human point of view.