Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling

We study the problem of 3D object generation. We propose a novel framework,namely 3D Generative Adversarial Network (3D-GAN), which generates 3D objectsfrom a probabilistic space by leveraging recent advances in volumetricconvolutional networks and generative adversarial nets. The benefits of ourmodel are three-fold: first, the use of an adversarial criterion, instead oftraditional heuristic criteria, enables the generator to capture objectstructure implicitly and to synthesize high-quality 3D objects; second, thegenerator establishes a mapping from a low-dimensional probabilistic space tothe space of 3D objects, so that we can sample objects without a referenceimage or CAD models, and explore the 3D object manifold; third, the adversarialdiscriminator provides a powerful 3D shape descriptor which, learned withoutsupervision, has wide applications in 3D object recognition. Experimentsdemonstrate that our method generates high-quality 3D objects, and ourunsupervisedly learned features achieve impressive performance on 3D objectrecognition, comparable with those of supervised learning methods.