Pothole Detector (YOLO26m)

A YOLO26m detector finetuned to spot road potholes from dashcam and street-level footage. Part of the Model Garden collection of finetuned computer-vision models.

Class

Pothole

Intended use

Built for road-condition surveys: pair it with a tracker and a GPS track to turn a dashcam drive into a geotagged map of pavement defects. Bounding boxes can be refined into tight masks with a promptable segmenter (e.g. SAM2) for area estimation.

Training

Base: YOLO26m (about 21.8M params)
Data: pothole.yaml
Epochs: 30
Image size: 512
Runtime: Apple Silicon (MPS)

Usage

from ultralytics import YOLO

model = YOLO("best.pt")
results = model("dashcam.mp4", stream=True)
for r in results:
    r.plot()

Limitations

Trained on street-level views; performance drops on aerial or very low-light footage. Detection is per-frame, so dedupe across frames before counting unique potholes. A research tool, not a certified road-safety inspection system.

Downloads last month: -