Dataset, code and run records for: Coarse-to-fine dual-branch inspection of bridge component defects using YOLO26 and DINOv3 (BD5/BS4 benchmarks derived from dacl10k)

Supplementary data and code for the manuscript Coarse-to-fine dual-branch inspection of bridge component defects using YOLO26 and DINOv3. Contents: (1) BD5, a five-class bridge-defect detection benchmark (crack, exposed rebar, spalling, rust, efflorescence/weathering; 7,910 images; train 6,000 / val 935 / test 975; YOLO format) and BS4, a four-class segmentation benchmark (background, crack, concrete damage, exposed rebar; 11,500 native-resolution 512x512 tiles), both derived from the public dacl10k dataset (CC BY 4.0); (2) conversion scripts, training and evaluation code of YOLO26-BD (strip-aware multi-scale attention, aspect-bounded Shape-IoU) and DV3-DRSeg (DINOv3 ViT-S/16 encoder with a dual-resolution decoder and decoupled dual gate) and of all baselines; (3) the records of all 88 training runs reported in the paper (metrics, curves, confusion matrices, gate statistics), latency/WCET/SAHI/area evaluations and the paper figures; (4) selected trained weights. DeepCrack is not redistributed (see README for the download link and split). The data zips are split by subset and merge into one folder when unzipped; see README.md.