CornerDYO: An Efficient Corner–Edge Reasoning Framework for Structured Building Vectorization From Remote-Sensing Images

Abstract

Structured building vectorization from high-resolution remote-sensing imagery provides essential geometric representations for urban mapping, digital twin construction, and geographic information systems (GIS). However, existing structured reconstruction approaches may still suffer from incomplete vertex localization, unstable edge connectivity, and increased computational demands when processing complex urban scenes. This article presents CornerDYO, an efficient corner–edge reasoning framework for structured building vectorization from remote-sensing images. The proposed framework integrates high-resolution feature enhancement, bidirectional feature fusion, context refinement, and corner-aware attention to improve geometric representation and structural connectivity. Experiments are conducted on the Outdoor Architecture Reconstruction benchmark introduced in previous structured reconstruction studies, containing 2,001 satellite image patches from Paris, Las Vegas, and Atlanta. Under the controlled evaluation setting, CornerDYO achieves 98.5% precision, 95.0% recall, and 96.7% F1-score for corner detection at $512 \times 512$ resolution, outperforming HEAT by 8.0 percentage points in F1-score. For edge detection, CornerDYO improves the F1-score from 79.7% to 83.8%. Moreover, under the same 500-epoch training schedule, CornerDYO reduces the training time from 6 h 28 min to 2 h 22 min. These results demonstrate that CornerDYO provides an accurate structural primitive extraction front end for remote-sensing building vectorization, achieving improved geometric prediction performance with reduced training time under the same experimental setting.