Interactive LSTM-Based Design Support in a Sketching Tool for the Architectural Domain

Johannes Bayer, Syed Saqib Bukhari, Andreas Dengel

In: International Conference on Pattern Recognition Applications and Methods. International Conference on Pattern Recognition Applications and Methods (ICPRAM-18) January 16-18 Funchal Portugal Scitepress 2018.


While computerized tools for late design phases are well-established in the architectural domain, early design phases still lack widespread, automated solutions. During these phases, the actual concept of a building is developed in a creative process which is conducted manually nowadays. In this paper, we present a novel strategy that tackles the problem in a semi-automated way, where long short-term memories (LSTMs) are making suggestions for each design step based on the user’s existing concept. A design step could be for example the creation of connections between rooms given a list of rooms or the creation of room layouts given a graph of connected rooms. This results in a tightly interleaved interaction between the user and the LSTMs. We propose two approaches for creating LSTMs with this behavior. In the first approach, one LSTM is trained for each design step. In the other approach, suggestions for all design steps are made by a single LSTM. We evaluate these approaches against each other by testing their performance on a set of floor plans. Finally, we present the integration of the best performing approach in an existing sketching software, resulting in an auto-completion for floor plans, similar to text auto-completion in modern office software.

ICPRAM2018-Bayer.pdf (pdf, 633 KB)

Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence