<![CDATA[Pothole detection based on deep learning has achieved high detection accuracy; however, most existing studies evaluate models using the same dataset for both training and testing, providing limited evidence of robustness under unseen data distributions. This study investigated the cross-dataset generalization capability of YOLOv8n using three publicly available pothole datasets with different visual characteristics: the Multi-Weather Pothole Dataset (MWPD), the Jaygala dataset, and the Andrew dataset. The proposed framework evaluated same-dataset and cross-dataset detection performance, quantified robustness through generalization gap analysis, examined the influence of dataset characteristics, and assessed the effectiveness of multi-dataset pretraining. Experimental results showed that the Andrew dataset achieved the highest same-dataset performance (mAP@50 = 0.816) but also exhibited the largest generalization gap (0.246), indicating limited robustness across datasets. In contrast, multi-dataset pretraining reduced the generalization gap for MWPD from 0.093 to 0.048, demonstrating improved cross-dataset robustness, although the improvement was not consistent across all datasets. These findings indicate that same-dataset accuracy alone is insufficient for evaluating model robustness and that cross-dataset evaluation provides a more realistic assessment of deployment performance in heterogeneous road environments.]]>
Copyrights © 2026
