Complex layout generation for large-scale floor plans via deep edge-aware GNNs

Abstract

In architectural layout generation, deep learning techniques have advanced the residential generation in multiple scenarios. However, current approaches fail to extract complex graph features from large-scale layouts, neglecting large-scale global context. Additionally, the lack of robust, quantitative evaluation metrics for layouts hampers the objective comparison of different generative approaches. To address these issues, we propose a multi-scale applicable layout generation method based on deep edge-aware GNNs, stressing edge-specific and non-local spatial information. Next, we introduce quantitative metrics to assess layout quality, including room accessibility index and space property proportion, whose purpose is to establish layout standards in the computer-aided design field. Lastly, we create the Public Space Floor Plan Dataset (P-PLAN), a collection of 4,535 annotated layout samples designed to serve as a robust evaluation platform for large-scale layout models. We conducted extensive qualitative and quantitative experiments on the Residential Floor Plan Dataset (R-PLAN) and P-PLAN dataset to demonstrate the effectiveness of the proposed method. Notably, with the proposed evaluation metrics, our method significantly outperforms existing models in accessibility and diversity.