CRTGAN: Controllable Road Network Graphs Generation via Transformer based GAN

Automated road network generation is in high demand among various applications. With the surge of generative AI, many methods synthesize road network images or graphs based on deep generative models. However, image-based approaches suffer from modeling topological features, while graph-based approaches struggle to encode spatial information. Furthermore, both of them fail to generate road networks with the desired properties. To alleviate these problems, we propose a novel controllable road network graphs generation pipeline to generate road network graphs with expected properties by focusing on both spatial and topological features. Our generation process is founded on modified transformer-based GAN architectures, which enhances the modeling of global features in road network images, such as topology-consistency. The synthesized road network image from the generator is simultaneously evaluated by the discriminator and the reward network, in aspect of spatial and topological supervision respectively, to control the generation process. Furthermore, we design a differentiable image-to-graph extractor applied in the middle of the pipeline. Extensive experiments demonstrate that our proposed pipeline arise generating road network graphs performance with desired properties, achieving 85.43% in traffic convenience metric.