3D Vectorization and Generation using Generative Adversarial Networks

Abstract

Generating and understanding the 3D shape of objects in the world is a crucial step for many areas of robotics. Across object categories, shapes are used for classification. Within each category, fine shape details and textures contribute to successful manipulation. Existing generation methods usually rely on sketches and meshes, new objects generated by obtaining and merging patterns and components from the database. The major drawback of such techniques is that they cannot produce a complete 3D object from a 2D image. Given a 2D image of a chair taken from front view missing back legs, in its corresponding 3D object this information will not be present.The architecture of the proposed model, employees 3D Vectorization and Generation using Generative Adversarial Network, that forms 3D-objects by leveraging the probabilistic space taking advantages from novel developments in volumetric convolutional networks and generative adversarial nets. The main advantages of the proposed method are: It uses an adversarial model that is capable of implicitly getting the object formation and to produce quality 3d objects. A mapping of 3Dobject is learned from a low dimensional probabilistic space. The adversarial discriminator gives a compelling 3D shape descriptor.