Rédiger un PREreview

The automated detection of surface defects on roads can help make condition assessments and develop maintenance plans as it allows road images to be transformed into machine-readable and localized data. The current research sets up a methodology for object detection and checks its performance for the detection of three surface categories: potholes, cracks, and manholes. The authors make sure that this research is considered as an empirical case study using a standard YOLOv8s detector rather than trying to prove a new detection method. The authors run the notebook acquiring our Road Damage Dataset which consists of 2009 manually labeled images and 4737 boxes in YOLO format. They produce a tidy bounding box table to guarantee the proper auditing of the dataset and conduct the necessary exploratory data analysis considering class balance, normalized geometry of the bounding box, spatial distribution of instances, co-occurrences of classes, image resolution, and constructed samples. Since the source dataset contains no validation collection in the notebook, a random selection of 15% images amounting to 302 images was held out, the detector achieves an overall precision of 0.921, recall of 0.877, mAP@0.50 of 0.953, and mAP@0.50:0.95 of 0.652. Manhole detection is strongest in mAP@0.50 (0.974), while crack detection provides the highest recall (0.895). Potholes are comparatively more difficult, with recall of 0.828 and mAP@0.50:0.95 of 0.607. The experiment demonstrates that a compact, pretrained one-stage detector can provide strong baseline performance on this three-class road-damage dataset. The results also expose important research limitations: the holdout is derived from the same original collection, no independent external test set is used, and no detector baseline or ablation study is reported. These limitations motivate future work on external validation, class-imbalance handling, detector benchmarking, and deployment-oriented evaluation.