Case Study · Model Training & Fine-Tuning
Custom Computer Vision Model for Vehicle Damage Inspection
A custom-trained detection model that finds and grades dents, scratches and panel damage from walk-around photos, so trade-in inspections are faster, more consistent and easier to defend with customers.
The problem
Trade-in and used-vehicle inspections depend on the individual inspector. Two people can grade the same car differently, reconditioning estimates vary, and customers dispute valuations that aren't backed by clear evidence.
How the model was trained
Data collection
Gather walk-around photos from past inspections across vehicle types, colours, lighting and weather.
Annotation
Label damage type (dent, scratch, crack, chip, paint mismatch), panel location and severity, using an annotation guide agreed with senior inspectors.
Transfer learning
Start from a pretrained detection and segmentation backbone and train it on the labelled automotive data.
Robustness
Augment for reflections, glare, dirt and camera angles, and rebalance so rare damage types are still learned reliably.
Evaluation
Measure mean average precision (mAP) by damage type, recall by panel, and agreement with a panel of experienced inspectors, and review where the model and experts disagree.
Deployment
Export an optimised, quantised model that runs in the inspection app on the device. Results flow into the dealer management system with an annotated photo report.
Feedback loop
Inspector corrections are captured and added to the training set, so the model improves with every inspection.
KPIs we design for
Technology
Have a similar challenge?
Talk to our team about what this could look like for your business.