Exploring optimal ways to represent topological and spatial features of building designs in deep learning methods and applications for architecture

Viktor Eisenstadt, Hardik Arora, Christoph Ziegler, Jessica Bielski, Christoph Langenhan, Klaus-Dieter Althoff, Andreas Dengel

In: CAADRIA 2021 Proceedings. International Conference of the Association for Computer-Aided Architectural Design Research in Asia (CAADRIA-2021) March 29-April 1 Hong Kong/Virtual Hong Kong SAR of China Cumincad 2021.


The main aim of this research is to harness deep learning techniques to support architectural design problems in early design phases, for example, to enable auto-completion of unfinished designs. For this purpose, we investigate the possibilities offered by established deep learning libraries such as TensorFlow. In this paper, we address a core challenge that arises, namely the transformation of semantic building information into a tensor format that can be processed by the libraries. Specifically, we address the representation of information about room types of a building and type of connection between the respective rooms. We develop and discuss five formats. Results of an initial evaluation based on a classification task show that all formats are suitable for training deep learning networks. However, a clear winner could be determined as well, for which a maximum value of 98% for validation accuracy could be achieved.


Weitere Links

caadria2021_086.pdf (pdf, 956 KB )

Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence