Generative Models for Context Dependent Urban Planning

Abstract

Building footprints represent the total area of coverage of a physical building. Footprints can be used to give an overview of a planned developed area in the early stages of urban planning. This thesis investigates the possibility of training a generative AI model to generate building footprints in a designated area based on surrounding, already existing building footprints. Such a generative model would be a useful tool for architects in the early stages of urban planning, as it would allow the rapid generation of footprint suggestions in an area designated for development. Two image inpainting networks trained for general image reconstruction were fine-tuned using a dataset of building footprints to improve on the task of footprint generation. One of the networks was also modified and trained to be able to accept a desired density of footprints in the generated area as an additional input. Two different masking algorithms were used during training and evaluation: A simple square approach and a more sophisticated algorithm that finds and masks city blocks. FID and LPIPS were used to evaluate and compare the trained models. It was shown that image inpainting networks form a good basis for context dependent footprint generation and that fine-tuning improves performance on this task. Furthermore, it was demonstrated that an image inpainting network can be modified to accept and adhere to density requirements providing a proof-of-concept for other types of user guidance.