TamAko783/Unified_Road_Defect_FRDC_Pseudolabeled_v2

Unified Road Defect Dataset - FRDC Pseudo-labeled v2 (two-teacher distillation set)

The exact training set used to train YOLO26n_RDD_FRDC_Distilled_v2. Unlike v1 (single Co-DETR teacher), v2 follows FRDC more closely by using two teachers and Weighted Boxes Fusion (WBF) to combine their predictions.

Contents

• GT train: the 4-class (D00/D10/D20/D40) Unified Road Defect Dataset (RDD-2022 + UAV-PDD2023 + RoadDamageVision) with original ground-truth labels.
• Pseudo-labeled test: the unlabeled RDD-2022 test images, labeled by the WBF fusion of two teachers from WangFangjun/FRDC-RDD: Co-DETR (Swin-L) + RTMDet-x. Only images with a confident fused detection are kept (rddtest_*).
• val: the original held-out GT validation split (unchanged), identical to v1 for a fair A/B against the v1 model and the GT-only baseline.

Layout (Ultralytics YOLO, 4 classes)

data/train_a.tar.gz  data/train_b.tar.gz   # GT train + fused-pseudo test
data/val.tar.gz                            # held-out GT val
rdd_frdc.yaml                              # Ultralytics data config (nc=4)
samples/                                   # a few fused-pseudo-labeled test images

Files prefixed rddtest_ are the two-teacher fused pseudo-labeled RDD-test images.

Classes

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

Credits

• Base dataset: TamAko783/Unified_Road_Defect_Dataset (RDD-2022 — Arya et al.; UAV-PDD2023; RoadDamageVision — Silva Zendron & Leithardt, CC BY 4.0).
• Pseudo-label teachers: Co-DETR Swin-L + RTMDet-x, FRDC (Wang Fangjun et al.), ORDDC'2024 winner; fused with Weighted Boxes Fusion (Solovyev et al.).
• Verify each source's license before redistribution / commercial use.