A Discussion on an Urban Layout Workflow Utilizing Generative Adversarial Network (GAN) - With a focus on automatized labeling and dataset acquisition

Abstract

Deep Learning (DL) has recently gained widespread attention in the automation of urban layout processes. This study proposes a rule-based and Generative Adversarial Network (GAN) workflow to automatically select and label urban datasets to train customized GAN models for the generation of urban layout proposals. The developed workflow automatically collects and labels urban typology samples from open-source maps. Furthermore, it controls the results of the GAN process with labels and provides real-time urban layout suggestions based on a co-design process. The conducted case study shows that the average value of the GAN results, trained from an automatically generated dataset, meets the site's requirements. The developed co-design strategy allows the architect to control the GAN process and perform iterations on urban layouts. The research addresses the research gap in GAN applications in the field of urban design and planning. Many studies have demonstrated that training the (GAN) model by labeling enables machines to learn urban morphological features and urban layout logic. However, two research gaps remain: (1) The manual filtering of GAN urban sample datasets to fit site-specific design requirements is very time-consuming. (2) Without a suitable data labeling method, it is difficult to manage the GAN process in such a manner to facilitate the meeting of overriding design requirements.

Building-GNN: Exploring a co-design framework for generating controllable 3D building prototypes by graph and recurrent neural networks

“…Some researchers have attempted to extract 2D features from 3D models and reconstruct 3D voxel models using encoding-decoding techniques. For instance, Zhong et al (2022) generated 3D volumes by using GauGAN to learn the height information recorded in the graph color channel, while Zhang & Huang (2021) used styleGan to learn and generate continuous profile features, which were then reconstructed into 3D models. However, these encoding-decoding methods do not fully capture the 3D details of the building model, such as the detailed 3D model features in the middle of two consecutive profiles, which are ignored in Zhang's method (Zhang & Huang, 2021).…”

“…The DL learning 3D voxelized models is widely used in architectural modeling (Chang et al, 2021). However, previous attempts, such as 3D-GAN are difficult to reason out complex constraint graph relationships between each voxel unit, and the 3D results are difficult to control precisely (Li et al, 2019;Oussidi & Elhassouny, 2018;Zhong et al, 2022), as will be discussed in detail in later chapters. The generated 3D building may not conform to the constraints and preference input by the architect, resulting in significant modifications.…”