TamAko783/Unified_Road_Defect_FRDC_Pseudolabeled_v1plus

Unified Road Defect Dataset - FRDC Pseudo-labeled v1+ (Co-DETR, gap-completed)

An improved version of the v1 distillation set. v1 used Co-DETR pseudo-labels at conf >= 0.5, which dropped 4,316 test images entirely (no confident detection) and missed lower-confidence boxes. v1+ keeps every v1 Co-DETR box unchanged and only adds the boxes v1 was missing (non-overlapping detections recovered from Co-DETR on the dropped images + RTMDet), so v1's strong labels are preserved while coverage is completed.

Unlike v2 (which fused both teachers with Weighted Boxes Fusion and averaged/moved Co-DETR's boxes), v1+ is add-only — no existing Co-DETR box is modified.

Contents

• GT train (RDD-2022 + UAV-PDD2023 + RoadDamageVision, 4-class, original labels).
• Pseudo-labeled test: 5,081 images, 9,160 boxes = v1's 8,023 Co-DETR boxes (kept)
  • 1,137 added missing boxes (362 images v1 had nothing for).
• val: the original held-out GT val (unchanged), identical to base/v1/v2 for fair A/B.

Classes

0 D00 Longitudinal | 1 D10 Transverse | 2 D20 Alligator | 3 D40 Pothole

Credits

• Base dataset: TamAko783/Unified_Road_Defect_Dataset.
• Pseudo-label teachers: Co-DETR Swin-L (primary) + RTMDet-x (gap-fill), FRDC (Wang Fangjun et al.), ORDDC'2024 winner.