A multibranch and multiattention framework for floor plan segmentation

Automated floor plan analysis is crucial in architecture, urban planning, and interior design. Floor plan segmentation is a foundational step for tasks such as surface area estimation and 3D building reconstruction. Unlike vector‐based approaches, our method operates directly on raster images, enabling it to process diverse and real‐world floor plan scans without requiring clean vector files, thus making it more practical for widespread applications. However, automatic semantic segmentation of floor plan images faces unique challenges, including high inter‐class similarity, ambiguous room boundaries, and varying floor plan styles. While several prior works have explored floor plan segmentation using conventional CNN architectures, many struggle with feature entanglement between room types and boundaries, often requiring additional post‐processing to refine results. To address these issues, we introduce a novel Multi‐Branch and Multi‐Attention framework (M&M) for floor plan segmentation. Our method employs a dual‐branch architecture that explicitly separates room‐type segmentation from boundary recognition, effectively avoiding feature entanglement and enabling each branch to specialize in its respective task. Furthermore, we incorporate intra‐branch channel attention and cross‐branch positional attention mechanisms to enhance both the precision of boundary detection and the accuracy of semantic room classification, significantly improving robustness and performance across diverse architectural layouts. Through extensive experiments on the raster‐to‐vector (R2V) and R3D datasets, we demonstrate how our approach sets a new state‐of‐the‐art for floor plan segmentation, outperforming general‐purpose and specialized models alike.