PolarSym: Polar Geometry-aware Attention for CAD Floorplan Parsing

CAD plan parsing constitutes a fundamental task in Building Information Modeling (BIM), whose objective is to automatically recover architectural elements such as walls, doors, windows, and furniture from 2D engineering drawings. Although existing Transformer-based methods can capture global semantic dependencies through self-attention mechanisms, they primarily infer spatial relationships from semantic features without explicitly modeling the inherent geometric symmetry of building layouts. Consequently, these methods are vulnerable to incorrect correspondences when handling long-range matching and complex symmetric spatial configurations. To address these issues, this paper proposes a polar-coordinate geometry-aware attention framework named PolarSym for CAD plan parsing. The framework explicitly decouples architectural geometric relationships into two complementary factors—direction and distance—and models them separately. It enhances structural consistency through directional constraints and establishes long-range symmetric correspondences via distance constraints. Additionally, a dynamic gating mechanism is employed to achieve synergistic fusion of the two geometric information streams while preserving the core transformer architecture. This design significantly improves geometric modeling capabilities with minimal computational overhead. Experimental results on a public CAD plan parsing dataset demonstrate that PolarSym achieves a performance improvement of 1.73% PQ, 1.54% RQ, and 4.31% mIoU over the reproduced SymPoint V2 baseline under identical training configurations. Furthermore, PolarSym exhibits faster convergence and more stable optimization dynamics. Ablation studies further validate the complementary role of directional and distance modeling. The results indicate that PolarSym effectively enhances the geometric perception capability of transformers with low computational overhead, providing an efficient geometric modeling solution for CAD plan parsing tasks.