+01 · Challenge
The customer needed road damage detected automatically in their own imagery — cracks, potholes, alligator cracks and side-road damage.
+02 · Approach
- 01Designed a multi-user annotation production pipeline
- 02Annotated the customer’s data
- 03Prepared training, validation, test and benchmark datasets
- 04Built training-data extraction utilities and quality checks
- 05Tuned hyperparameters, trained, evaluated and benchmarked the model
+03 · Result
A trained and benchmarked detection model for four damage classes, delivered with the tooling to keep extending the training data.
