Self-Optimization for Parsing Floor Plans

Abstract

Floor plan vectorization is an emerging research area in geographic information science and computer vision. However, automated recognition of building elements remains a challenge. This work proposes a method that combines the advantages of classical graphics with deep learning. Specifically, a morphological template is introduced to optimize topological relations, enhance completeness, and suppress conflicts. Bezier curves are utilized to represent irregularity contributing to improving visual effects and experimental accuracy. Thus, the proposed method can be directly practiced to boost performance and correct pseudo-samples in self-training. Experiments demonstrate that the proposed method achieves a considerable improvement in CVC-FP and R2V benchmarks. Additionally, our approach outputs instances with consistent topology, enabling direct modeling into Industry Foundation Classes (IFC) or City Geography Markup Language (CityGML). Hopefully, this work can serve as a new baseline for further study.