A Comparative Study of Two Road Pothole Datasets for Instance Segmentation Using YOLO-Seg and K-Fold Cross-Validation

Abstract

Road potholes are a persistent form of pavement deterioration that threatens public safety and imposes recurring costs on transportation agencies. Vision-based instance segmentation offers a scalable and precise approach for delineating such defects, with YOLO-family models extended with segmentation heads achieving strong accuracy and inference efficiency for field deployment. Yet despite a growing number of public pothole datasets, the effect of dataset choice on segmentation performance has not been rigorously isolated: most studies use a single dataset or compare different models on different data, conflating architectural and data effects. This paper addresses that gap through a controlled comparison of two public datasets, which is the Pothole Image Segmentation Dataset of YOLOv8 (Dataset A, 780 images) and the Semantic Segmentation Dataset of Indian Roads (Dataset B, approximately 7,640 images), isolating dataset choice while holding every other variable constant. To evaluate model resilience under a strictly controlled limited data condition, 300 images are sampled from each dataset with a fixed seed (67) and processed through a unified mask-to-YOLO-polygon pipeline; both subsets train a YOLOv26n-seg model under identical Copy paste augmentation (mosaic = 1.0, copy_paste = 0.4), with generalization estimated via 5-fold cross-validation. Dataset B outperforms Dataset A on every metric: mean mask mAP@50– 95 of 0.5880 versus 0.2559 (a 2.3× gain) and an mAP@0.5 plateau near 0.88 versus 0.50, with a smaller train–validation loss gap (0.7575 versus 1.6752) and lower instantaneous GPU utilization (23.83% versus 33.75%). The advantage is consistent across all five folds, establishing that in a small-data regime, dataset quality which is visual diversity and annotation specificity, is the primary driver of segmentation performance.