Authors:
Sarun Poolkrajang
and
Anand Bhojan
Affiliation:
School of Computing, National University of Singapore, Singapore
Keyword(s):
Generative Adversarial Networks, Neural Networks, Computer Vision, City Generation.
Abstract:
City generation for video games is a resource and time-consuming task. With the increasing popularity of open-world games, studies on building virtual environments have become increasingly important for the game research community and industry. The game development team must engage in urban planning, designate important locations, create population assets, integrate game design, and assemble these elements into a cohesive-looking city. Based on our limited knowledge and survey, we are the first to propose a holistic approach that integrates all features in generating a city, including the natural features surrounding it. We employ a generative adversarial network architecture to create a realistic layout of an entire city from scratch. Subsequently, we utilize classical computer vision techniques to post-process the layout into separate features. The chosen model is a simple Convolutional GAN, trained on a modest dataset of 2x2 km² snippets from over two thousand cities around the wo
rld. Although the method is somewhat constrained by the resolution of the images, the results indicate that it can serve as a solid foundation for building realistic 3D cities.
(More)