2027
Lecture Notes in Computer Science, Scopus
Статьи в журналах
Vol 16790. Pp. 319-330.
Kostin D., Gribova V. CGAN Model for Building Layout Plans Generation. In: Ronzhin, A., Gribova, V., Meshcheryakov, R. (eds) Interactive Collaborative Robotics. ICR 2026. Lecture Notes in Computer Science. 2027. Vol 16790. Pp. 319-330. Springer, Cham. https://doi.org/10.1007/978-3-032-34387-1_23.
The task of layout planning is a fundamental challenge in developing site masterplans, governed by spatial and architectural constraints. Automating plan generation can greatly reduce the effort required to produce design alternatives. In addition, multiple design variants can be used as digital twins of real-city morphologies for training navigation in robotics. While shape grammars and genetic algorithms have been used for this purpose, they require manual adjustment of transformation rules and optimization criteria for each new case. In contrast, deep learning methods can identify patterns and generate designs based on existing developments. Conditional Generative Adversarial Networks (CGANs) allow constraints to be incorporated by using label maps as input. In this work, the Pix2PixHD model—featuring one generator and two discriminators—is applied to the problem. Building layout plans are generated for sites in Vladivostok, Saint Petersburg, Sochi, Seoul, and San Francisco, and the results are compared using numerical image similarity metrics. The primary metrics are Fréchet Inception Distance (FID) and Structural Similarity Index (SSIM), chosen for their ability to compare distributions. The key advantages and limitations of the model are discussed. The paper concludes with an assessment of the applicability of this deep learning approach to solving the layout planning problem in architectural design. Directions for future research have been identified, including the use of hybrid models as well as addressing the need for training on giant datasets.