Robust Attributed Adjacency Graph Extraction Using Floor Plan Images

Abstract

Architectural design solutions are intrinsically structured information with a broad range of interdependent scopes. Compared to conventional 2D Euclidean data such as orthographic drawings and perspectives, non-Euclidean data (e.g., attributed adjacency graphs) can be more effective and accurate for representing 3D architectural design information, which can be useful for numerous design tasks such as spatial analysis and reasoning, and practical applications such as floor plan parsing and generation. Thus, getting access to a matching attributed adjacency graph dataset of architectural design becomes a necessity. However, the task of conveniently acquiring attributed adjacency graphs from existing architectural design solutions still remains an open challenge. To this end, this project leverages state-of-the-art image segmentation techniques using an ensemble learning scheme and proposes an end-to-end framework to efficiently extract attributed adjacency graphs from floor plan images with diverse styles and varied levels of complexity, aiming at addressing generalization issues of existing approaches. The proposed graph extraction framework can be used as an innovative tool for advancing design research infrastructure, with which we construct a large-scale attributed adjacency graph dataset of architectural design using floor plan images retrieved in bulk. We have open sourced our code and dataset.