Answers are grounded in this paper and cite their sources.
Comparative evaluation of YOLO variants improves detection accuracy in buried manhole cover applications, suggesting optimal model selection for infrastructure inspection.
Urban paving and resurfacing operations frequently bury utility access points, complicating infrastructure maintenance and emergency response. Although ground‐penetrating radar (GPR) can be used to locate buried manholes, manual radargram interpretation requires specialised expertise and is time‐consuming and subjective. This paper introduces a systematic comparative evaluation of three YOLO (You Only Look Once) variants of different architectural generations, including YOLOv5s (anchor‐based), YOLOv8s (anchor‐free with decoupled heads), and YOLOv11s (transformer‐enhanced). We trained and validated these models using 9‐fold cross‐validation on 54 controlled radargrams obtained at 1000 MHz under systematically varied conditions: three surface types (asphalt, concrete, sand), six burial depths (0.05–0.75 m), and three compaction levels (C0, C5, C10). The results show that YOLOv5s showed the best performance under sparse data conditions, with accuracy = 0.7865, precision = 0.8198, recall = 0.7533, mAP@0.5 = 0.7950 and the lowest cross‐fold variability (RMSE = 0.0725). YOLOv8s achieved higher mAP@0.5:0.95 (0.3711), indicating good generalisation on strict IoU thresholds, but YOLOv11s showed high instability (RMSE = 0.2200), indicating overfitting. Paired t‐tests showed no statistically significant difference between architectures at α = 0.05, but practical stability differences that are deployment‐relevant were prominent. Detection accuracy degraded systematically at high compaction (C10) and at greater burial depths, consistent with electromagnetic attenuation in dense media. YOLOv5s also offered the shortest inference time, which supports near‐real‐time deployment scenarios. A prototype of the interface based on the LINE is used to demonstrate the technical feasibility of field workflows. These results suggest that the lighter YOLO architecture provides better stability and efficiency for GPR‐based utility detection when training data is limited. This can guide the model selection for infrastructure inspection applications.
No takes yet. Share an insight, caveat, or question.
Panyavaraporn et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: