Deployment-oriented benchmark for automated UAV-based facade defect detection

Abstract

Integrating computer vision with UAVs can automate parts of building facade inspection while reducing labor and access risks. Applying UAV-based defect detection to regional multi-building screening, however, requires planning of coverage-derived data construction, task definition and labeling effort, detector selection, and recurring-error mitigation. This paper presents BFD-UAV2K, a benchmark comprising 2000 full-frame images collected from 42 reinforced-concrete buildings, with three-class defect annotations and building-disjoint data partitions. Eight detectors are evaluated across five training-set sizes with recorded-and-extrapolated annotation effort and standardized GPU inference efficiency. One-stage models provide favorable throughput and low-data screening behavior, whereas query-based detectors trade higher recall for greater latency and false-alarm burden in selected classes. Multi-scale training, hard-negative mining, and geometry-aware cropping are evaluated after baseline diagnosis; their effects are nonuniform across classes and architectures. The findings support construction automation by linking coverage-oriented UAV acquisition with annotation planning, detector selection, and targeted optimization for multi-building facade screening.