情報処理学会 第80回全国大会講演要旨

This paper introduces a novel approach for generating 3D shapes of objects by employing generative adversarial networks(GANs). While GANs have been used in this task, previous work focused on learning a mapping from random latent vectors to unpredictable 3D shapes. We address the issue of generating 3D shapes of our interest by embedding category information into latent space. Our model learns a representation of this information and generates both diverse and discriminable 3D shapes comparable to real data, while model complexity is not increased. Furthermore, we extend the model to utilize more detailed description of 3D shapes besides category information, resulting in more accurate generated shapes.