Sketch to Building: Architecture Image Translation Based on GAN

Abstract

Image-to-image translation with Generative Adversarial Networks has been extensively studied recently. However, it remains still very challenging for sketch-to-image translation owing to large domain gap and information asymmetry problems. In this paper, we introduce a novel two-step method as well as designing an interactive system which can generate different styles of architecture images from a hand-drawn sketch. In our framework, either an architecture sketch can be drawn and then translated into a color image, or an architecture photo can be loaded, followed by sketch generation, modification, and translation. Style images can also be manually selected for different output styles. Experimental results show that our method has better performance on both image quality and diversity than the other competitive methods.