Simplifying the bipartite matching topology recovery for vectorized building footprint extraction

Extracting vectorized building footprints from high-resolution imagery requires precise corner localization, robust boundary connectivity, and the generation of spatially ordered topological structures that faithfully encompasses the building area. These interdependent requirements make current approaches complex, which typically rely on cascaded architectures combining heterogeneous neural networks for feature extraction and topological recovering separately. This paper proposes ABCNet, a unified CNN architecture that simultaneously extracts instance-level areas, boundaries, and corners while robustly recovering ordered topological sequences for building footprints. Specifically, we augment a standard YOLO-based instance segmentation network with dual detection heads to generate per-building instance mask scores for areas, boundaries, and corners. The initial exterior contour is derived from the instance area mask. To refine the footprint geometry beyond smoothed contour approximations, we establish geometrically ordered corner connections through a bipartite matching matrix. Each entry in this square matrix encodes the interpolated boundary probability score between two corners indexed by its row and column. Ordered relationships are estimated using path-length differences relative to the initial contour, computed from parameterized distances of orthogonally projected points along the contour. The vertex sequence direction (clockwise/counterclockwise) is resolved via the triangle inequality criterion. Finally, the topological sequence is recovered through linear-sum-assignment optimization applied to the bipartite matching matrix, giving the vectorized building. Experimental evaluation on two vectorized building extraction benchmarks demonstrates that the proposed method, despite its architectural simplicity, surpasses state-of-the-art approaches by significant margins—achieving 8.6% and 2.5% higher AP (Average Precision) under in-distribution and out-of-distribution scenarios, respectively; with a 3% improvement on human-annotated out-of-distribution vectorization benchmarks. Visual comparisons further reveal the method’s robustness, maintaining coherent degradation patterns even when geometric primitives are misdetected under challenging imaging conditions. The simple architecture inherently preserves design flexibility, enabling future enhancements to area, boundary, and corner detection precision through dedicated module refinements. • Resolves the inherent ambiguity of simple MLP-based affinity modeling for corner pairs. • Intermediate features (areas, boundaries, corners) from feature extraction enable explicit vectorized modeling independent of the backbone architecture. • Establishing unambiguous direction and accurate boundary relationships is crucial for topological reconstruction via optimal assignment matrix. • Running-length-based corner ordering supports orientation resolution through triangle inequality principle.