From instance segmentation to physical quantification: High-resolution UAV-based dataset for façade defect assessment

High-rise building façade inspection is essential but hazardous and labor-intensive. Although UAV imaging and defect segmentation methods can recognize defects, most studies stop at 2D masks and lack the scale and instance structure needed for grade quantification. Public façade datasets remain scarce, often low-resolution, crack-only, and annotated for semantic segmentation. This paper introduces CUBIT-InSeg, a high-resolution UAV building façade defect instance segmentation dataset. It contains 6996 images at 4800 × 3200 with 62,187 annotated instances of two safety-critical defects, cracks and spalling, in realistic multi-instance scenes. Over 80 models were benchmarked under a unified protocol, and transferability was assessed via zero-shot testing on two cross-domain datasets. A deployment workflow was further demonstrated that registers defects to an as-is digital twin and converts segmentation masks into metric measures (e.g., crack width/length, spalling area) for standard-aligned severity evaluation. In real-world validation, the workflow based on CUBIT-InSeg dataset enables reproducible research toward scalable, quantitative façade condition assessment.