Generation of Figures with Controllable Posture using Ss-InfoGAN

摘要

With the increasing interest of social media, individual content production has become widespread. However, producing quality content will require much time and skills for individuals. On the other hand, recent generative adversarial networks (GAN) can easily generate images by learning. We aim to support or assist the video production and lower the threshold of content production by making the output results of the deep generation model more controllable. However, in conventional GANs, the correspondence between input and output was not easy for humans to interpret. As a premise, one interpretable example is that each input corresponds to each element of the output image. Therefore, in this research, we aim to control poses that are easy to interpret for images generated from 3D models of people using Ss-InfoGAN. Each input of Ss-InfoGAN is associated with the inclination of each joint, or only one input is moved. Experiments are conducted to check whether only the expected joint changes, and succeeded in actually associating the input with the joint state.