Pixel-Level Pavement Distress Assessment Using Instance Segmentation

Researchers used Mask R-CNN in Detectron2 to assess pavement distress on a custom roadway image dataset.

The best model, ResNet-101 FPN, achieved strong precision, recall, and F1, and its crack-area estimates closely matched ground truth.

A retrained YOLO baseline performed much worse, showing that instance segmentation is better than bounding-box detection for thin, irregular defects.

The results support pixel-level crack localization for more accurate roadway damage quantification and maintenance planning.