RUB » Bibliotheksportal

Fadi Khamam, Sven Schneider

• Floor plan design is a complex task that relies on human interaction, intuition, and appropriate tools to configure space that meets the clients’ needs. While automation holds the potential to simplify this process, it faces challenges in adhering to diverse rules, performance criteria, and boundary conditions. In the past decade, many approaches have been developed to automate floor plan generation, including rule-based and machine learning-based methods. Our study focuses on the usage of deep learning for designing residential floor plans and - in particular - to control the generation process by user interaction. We apply Conditional Generative Adversarial Networks (cGANs), namely the pix2pix model, for converting a bubble diagram representing different room topologies and their positioning within a given apartment boundary into a labeled layout that adheres to geometrical rules and constraints, such as correct proportions, area, and boundary lines. Via the bubble diagram, it provides the ability to interact with the proposed model giving the user more control over the generated output. In this paper, we present the process of dataset preparation, cleaning, annotation, and model training. The generated results of the model are evaluated based on the following criteria: room count, connectivity, room partitioning, and response to user interaction.