Road Damage Detection (YOLO11)

Build a baseline road-damage detector (mainly potholes) with honestly measured scores, plus a small decision layer on a dashcam video demo. The limits are stated on purpose: not field-tested on Indian roads, not real time, and no accuracy claim on video.

Trained a YOLO11n baseline, then YOLO11s, on the RDD2022 India data (4 classes: longitudinal, transverse and alligator cracks, and potholes). The India-only model barely worked on UK pothole photos (Kent dataset), so I fine-tuned on India + Kent. A confidence cutoff of 0.35 was chosen on validation, because a wrong box seemed worse than a missed pothole. The video demo adds ByteTrack tracking, an editable road-area filter, duplicate-ID merging and a transparent 0-100 severity score.

YOLO11s on the India test set: mAP50 0.332 and pothole AP50 0.453. The default India + Kent model scores pothole AP50 0.413 on the India test and 0.218 on a held-out Kent test, up from 0.020 before fine-tuning. At the 0.35 cutoff about 4 in 10 potholes are found (recall 0.366 on India, 0.305 on Kent). Single runs, and the scores are slightly optimistic because the split comes from the Kaggle upload.

A domain gap: the India-only model scored 0.020 pothole AP50 on UK photos. Early false alarms on a brick wall, kerbs and road markings had confidences of 0.41-0.62, above the cutoff, so I added a road-area filter instead of just raising the cutoff. Transverse cracks can't be judged (only 7 test boxes). The video is UK footage, with no confirmed pothole count or false-alarm rate, and one clear false alarm on a white road marking.

Only claim what is built, a lesson carried over from the AnnaDaan demo. A road-area filter beats a higher cutoff when false alarms score high, and a test photo that was also in the training data makes a check unfair (leakage).