Floor Plan Generation and Auto Completion Based on Recurrent Neural Networks

Johannes Bayer, Syed Saqib Bukhari, Andreas Dengel

In: The 12th IAPR International Workshop on Graphics Recognition (GREC 2017), Kyoto Japan, November 2017.. IAPR International Workshop on Graphics Recognition (GREC-2017) November 9-10 Kyoto Japan IEEE 2017.


During early design phases, the architect's task is to develop a floor plan layout from a high level description. This process is usually conducted manually nowadays in an iterative manner. In order to assist the architect with repetitive tasks during the individual design steps, we trained a recurrent neural network to mimic the architect's behavior. Our approach is based on sequences that recreate the user's behavior and that we generated from simple floor plans. By utilizing a dedicated inferencing mechanism, we are able to implement the generation of different design steps and tasks using a single LSTM model. We compare two different types of sequencing approaches by calculating their errors on a test set for a selected design step and evaluating the results qualitatively. While the current performance still needs to be improved for productive use, our dedicated inference mechanism shows a functional behavior.

Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence