Towards the rule-based synthesis of realistic floor plan images

Building Information Modeling (BIM) relies on 3D models of buildings, which are often time-consuming and expensive to acquire. To address this challenge, deep learning-based computer vision techniques can be used to generate 3D models from 2D drawings automatically. However, training these deep learning models requires large labeled data sets, and creating custom data sets involves tedious manual labor. To overcome this limitation, we propose an approach to generate synthetic labeled floor plan images for training deep learning models. The method produces floor plans with dimension lines, textual room information, and object detection labels in the Common Objects in Context (COCO) format. Potential future improvements include enhancing drawing style diversity and incorporating additional components such as furniture and technical building equipment