Street Layout Design via Conditional Adversarial Learning
share
Designing high-quality urban street layouts has long been in high demand, but entangles notable challenges. Conventional methods based on deep generative models are yet to fill the gap on integrating both natural and socioeconomic factors in the design loop. In this paper, we propose a novel urban street layout design method based on conditional adversarial learning. Specifically, a conditional generative adversarial network trained on a real-world dataset synthesizes street layout images from the feature map, into which an autoencoder fuses a set of natural and socioeconomic data for a region of interest; The following extraction module generates high-quality street layout graphs corresponding to the synthesized images. Experiments and evaluations suggest that the proposed method outputs various urban street layouts that are visually and structurally alike their real-world counterparts, which can be used to support the creation of high-quality urban virtual environments.
READ FULL TEXT
