STAYCHARTED BLOG · MACHINE LEARNING · VEHICLE DAMAGE
Machine Learning for Vehicle Damage Classification: What Three Photo Tests Show
Machine learning for vehicle damage classification, tested on car photos: 86.4% on damage, 89.9% on parts, 70.2% on damage type, and where people still check.
Why machine learning for vehicle damage classification works on your own photos
Keyword rules do nothing for a photo, and a general AI model that describes a picture does not know your categories. It can say "the bumper looks dented," but it does not know where your team draws the line between breakage and crushed, or when a mark counts as a scratch and when it counts as a paint chip. Those rules are rarely written down. They live in years of decisions your claims and inspection staff have already made.
Supervised machine learning turns those decisions into a model. You give it photos with the labels your team assigned, it learns the boundaries, and it is tested on photos it never saw before you rely on it. That gives you:
- Your categories, with your boundaries. The model learns the split your team uses, including the close calls.
- A published model version. Apply the learned categories across inspection batches and record which version produced each answer.
- A confidence you can check. The held-out test shows how often answers above each cutoff were right, so you can choose how much goes straight through.
- A model you can improve. Photos a reviewer corrects become training examples for the next version.
Car damage detection machine learning: the benchmark results
StayCharted AI Model Trainer (AMT) trained two image models on each test and scored them on held-out photos they never saw in training. Both models learned from the same labeled pictures.
| StayChartedAI Model Trainer · model comparison | ||||
|---|---|---|---|---|
| Test | Categories | Training / test photos | StayCharted AI Image Classifier Model | StayCharted Dedicated AI Image Model |
| Damage | 6 | 1,655 / 552 | 76.6% | 86.4% |
| Exterior part | 21 | 1,890 / 840 | 85.5% | 89.9% |
| Damage type | 8 | 1,392 / 449 | 64.1% | 70.2% |
The AI Image Classifier Model uses a pre-trained vision model to turn each photo into numbers, and a small classifier learns your categories from them. It trained in 0.3 minutes on every test. The Dedicated AI Image Model fine-tunes the AI model itself on your photos, which took 33 to 54 minutes here, including time waiting for a GPU. The Dedicated model was more accurate on all three tests.
Damage: the trained model gains most on the subtle split
The damage test breaks down into three questions. The gap between the two models is very different from one question to the next.
- Front or rear: 100% for both models.
- Damaged or not: 88.6% for the AI Image Classifier Model, 94.0% for the Dedicated AI Image Model.
- Breakage or crushed (damaged cars only): 79.7% and 87.8%.
The easy question needed no extra model at all. The gain came on breakage versus crushed, which is where most mistakes were and, as the benchmark notes, where people disagree most too. That pattern is the main reason to train a dedicated model: the closer two categories are, the more your own labeled examples matter.
Car part recognition from photos
The Dedicated model identified the right part out of 21 in 89.9% of 840 test pictures. License plates, mirrors, tail lights and rocker panels were 97.5% or better, and every part except the front wheel was 80% or better. The front wheel was found only 52.5% of the time, because front and rear wheels look alike when cut out of the photo. Treating them as one "wheel" part raised overall accuracy to 93.7%.
The practical lesson is to define categories around what the photo can show. If your process truly needs front versus rear wheel, give the model more context, such as the whole car or a text column with the adjuster's note.
Damage type classification is the hardest, at 70.2%
Damage type was the weakest result, and it is worth being plain about it. The Dedicated model scored 70.2% overall across eight types and 77.7% on the four common ones (broken part, dent, missing part, scratch). The four rare types (corrosion, cracked, flaking, paint chip) had only 41 to 67 training pictures each, and both models mostly missed them: 19.0% for the Dedicated model and 13.8% for the AI Image Classifier Model. The rare types were mostly called scratch.
A model learns the damage you show it. If corrosion or flaking matters to your business, label more examples of them before trusting the model on those types, and send them to a person until then. Our guide on how many examples you need to train a classifier covers this in more detail.
Vehicle damage image classification with a person in the loop
Each answer comes with a confidence from 0 to 1. Answers below your cutoff go to the Review Queue, where an adjuster or advisor confirms them. The default cutoff is 0.5 for the AI Image Classifier Model and 0.9 for the Dedicated AI Image Model.
| StayChartedAI Model Trainer · model comparison | |||
|---|---|---|---|
| Test | Model | Sent to a person (default cutoff) | Projected after review |
| Damage | AI Image Classifier Model | 11.6% | 83.5% |
| Damage | Dedicated AI Image Model | 23.0% | 95.8% |
| Part | AI Image Classifier Model | 24.3% | 96.1% |
| Part | Dedicated AI Image Model | 15.5% | 96.8% |
| Damage type | AI Image Classifier Model | 36.1% | 84.4% |
| Damage type | Dedicated AI Image Model | 50.8% | 93.5% |
On damaged or not, 77% of photos were above the Dedicated model's cutoff, and 98.4% of those were right. Four of 552 were damaged cars labeled undamaged above the cutoff, which is the mistake an insurer cares about most. On parts, 85% of pictures were above the cutoff and 96.2% of those were right. Damage type sent 50.8% of answers to a person, so plan for real review effort there.
Two points stand out. First, the "projected after review" column assumes a reviewer corrects every flagged mistake without adding errors; it is a projection, not a measured result. Second, the AI Image Classifier Model sent fewer damage photos to review (11.6%) yet projects only 83.5%, which means more of its mistakes sat among answers it was confident about. Check the confidence report before you set a cutoff. Our post on confidence scores and human review explains how to choose one.
AI car damage assessment from photos: where it fits
- Insurers and claims teams: sort incoming claim photos into damaged or not, and front or rear. Confident answers route straight to the right queue and uncertain ones go to an adjuster. See insurance claims triage.
- Fleet and rental operators: tag return and check-in photos by damage and part, with uncertain photos sent to a person before a charge or repair order.
- Body shops and dealer service: tag part and damage type before an advisor opens the job. Part recognition is strong, but damage type needed a person on about half the photos.
- Vehicle makers: group returned-part and field photos by part and damage type, starting with the part model. See visual inspection.
When you don't need machine learning
Not every photo task needs a trained model. If you handle a few photos a week, a person is faster than any setup. If your categories are broad and visually obvious, the AI Image Classifier Model may be enough on its own; front or rear was 100% for both models. Train the Dedicated AI Image Model where categories are close and a mistake costs money, such as missed damage on a claim. If you have no labeled photos yet, label a sample first, because the model can only learn decisions that already exist.
How to start a vehicle damage classification model
- Export photos your team already labeled, with the category in a column. Upload them, or point AMT at S3-compatible storage for large photo sets.
- Train an AI Image Classifier Model first. It needs the Vision add-on on Essentials or Business and trains in minutes.
- Read the held-out report: accuracy overall and per category, the confusions, and the share of answers above each cutoff with how many were right.
- Train a Dedicated AI Image Model if close categories matter. It needs Business with Vision. Compare the two on the mistakes that matter to you.
- Set a cutoff and fill new photos through file uploads, the API or the Power Automate connector. Answers below the cutoff go to the Review Queue, and reviewers' answers train the next version.
Your data stays in your organization's own database area and is used only for your models. See image classification, the product overview and pricing.
These results come from one run per model on public datasets, with test cut-outs that are not customer photos, and human review was not measured. Your own photos will score differently, so validate on held-out photos from your claims or inspections before relying on a model.
The Review Queue requires Essentials or Business. Retrain and publish a new version to use reviewer corrections in future answers.
Common questions
What is machine learning for vehicle damage classification?
It is training a model on car photos your team has already labeled (damaged or not, part, damage type) so it sorts new photos into the same categories. The model is tested on photos it never saw and gives each answer a confidence, so uncertain photos go to a person.
How accurate is AI car damage detection from photos?
In our benchmark the Dedicated AI Image Model scored 94.0% on damaged or not and 86.4% across six damage categories on 552 held-out photos. On damaged or not at the default cutoff, 77% of photos were above it and 98.4% of those were right.
Can machine learning classify the type of car damage?
Partly. Damage type classification was the hardest test: 70.2% overall across eight types and 77.7% on the four common types. Rare types such as corrosion and flaking had 41 to 67 training pictures each and were mostly missed, so they need more examples and human review.
Can AI recognize car parts from photos?
Yes. The Dedicated AI Image Model named the right part out of 21 in 89.9% of 840 test pictures, and 93.7% when front and rear wheels were treated as one part. License plates, mirrors, tail lights and rocker panels were 97.5% or better.
How many photos do I need to train a car damage classification model?
It depends on how close your categories are. In the part test, 90 training pictures per part gave 89.9% accuracy, while damage types with only 41 to 67 pictures each were mostly missed. Start with the photos you have and add examples where the report shows confusion.
Which StayCharted plan do I need for vehicle damage image classification?
Image models need the Vision add-on. The AI Image Classifier Model works on Essentials or Business with Vision, and the Dedicated AI Image Model needs Business with Vision. See pricing for current plans.