Two-stage Floor Plan Recognition with a Network Employing Attention Mechanism

The floor plan recognition task often requires architects to craft with their professional skills and tools. To solve this problem, this paper introduces a new two-staged approach for floor plan recognition. In the initial stage, rough segmentation of room regions is accomplished. Utilizing an attention mechanism guided by room boundaries, our network is directed to focus on the relationship between house boundaries and room regions, resulting in pixel-level segmentation for room regions. In the second stage, we integrate the rough segmentation result with the boundary vectors and architectural symbols. The merged information undergoes a weighted scoring mechanism, then derives a unified semantic label for all pixels within a boundary closed-loop region. In summary, our approach not only achieves the room-level semantic segmentation, but also captures the interplay among boundaries, architectural symbols, and room regions, thus attains a comprehensive understanding of floor plans.