Data Augmentation Strategy Evaluation for Pothole Instance Segmentation Using YOLOv26n-seg on Resource-Constrained Hardware

Abstract

Pothole and road surface deterioration pose safety risks that demand automated road inspection systems. This paper investigates five data augmentation strategies, namely Baseline, Copy-Paste, MixUp, CutMix, and Hybrid (Copy-Paste + CutMix), to improve instance segmentation performance for pothole detection using YOLOv26n-seg on resource-constrained hardware. A dataset of 300 manually annotated road images was evaluated via 5-fold cross-validation. Results show CutMix achieved the highest detection rate (mAP@50 = 0.5350, Precision = 0.6519), Copy-Paste achieved the best localization (mAP@50-95 = 0.2647), MixUp recorded the smallest loss gap (0.1153), and Baseline achieved the highest energy efficiency (22.58 mAP/kWh).