AUTOMATIC GENERATION OF REALISTIC CITY IMAGES FROM RARE DATASET USING GAN ENHANCED WITH TRANSFER LEARNING

Abstract

There has been a growing demand to strengthen existing disaster prevention education tobe prepared for the huge tsunami expected to occur in the near future. Virtual reality, which allows people to virtually experience natural disasters, has a strong potential in fostering disaster awareness among citizens. However, it requires enormous human and time resources to map the texture of structures to urban area-imitating virtual space. On the other hand, pix2pixHD proposed by Wang et al. can generate high-resolution synthetic images by learning from reference images, label data, and object boundary data. In this study, we applied pix2pixHD and transfer learning, which diverts networks trained on other similar domains, to verify texture mapping of urban areas in Japan from a limited set of image data.