Car photos for claims, service and OEMs: damage, damage type and part, measured

Insurers, body shops, dealer service lanes and vehicle makers all look at the same photos and ask the same questions: is the car damaged, what kind of damage, and on which part? We trained both of StayCharted AI Model Trainer's picture models on public car photos and scored them on photos they had never seen.

• Damaged or not (2,207 photos, front and rear of cars):
  • The Dedicated AI Image Model was right on 94.0% of 552 test photos.
  • At its default review cutoff, 77% of photos were above the review cutoff, and 98.4% of those were right about damaged or not.
  • 4 of 552 were damaged cars labeled undamaged above the review cutoff.
  • If a person corrected every flagged error in the other 23%: projected 98.7% accuracy on damaged or not.
• Which exterior part (21 parts, cut out of whole-car photos):
  • 89.9% exactly right; 93.7% if front and back are not told apart.
  • 85% of pictures were above the review cutoff, and 96.2% of those were right.
• What kind of damage (8 types):
  • The hard one. 70.2% overall, and 77.7% on the four common types (dent, scratch, broken part, missing part).
  • Four rare types (corrosion, cracked, flaking, paint chip) had only 41 to 67 training pictures each, and the models mostly missed them. This is the lesson, not a footnote: a model learns the damage you show it.
  • Here the review step carried the result: 50.8% were flagged for review, giving projected 93.5% accuracy if every flagged error were corrected.

The pattern for an AI agent handling claim or intake photos: the clear cases go straight through, and the uncertain ones go to an adjuster or service advisor. Reviewed corrections can become training examples for the next version.

How the test sets were chosen:

• Damage: 2,300 published photos. We removed 83 exact or near copies, and dropped 10 more that appeared under two different categories, since their label is in doubt. Then 25% of each category was held out, by a fixed shuffle.
• Part and damage type: the Humans in the Loop photos come with outlines drawn around each part and each damaged area. We cut each outline out as its own picture, with a 15% margin. 25% of the original photos were held out before any cutting, so no test picture comes from a photo the model trained on. Cut-outs under 48 pixels on a side were skipped.
• All pictures were resized so the longest side is at most 512 pixels, which is what AMT keeps and what both models read. Test pictures had neutral file names.

• AI Image Classifier Model (Vision add-on, Essentials and Business):
  • a pre-trained vision AI turns each photo into a numerical representation (embedding);
  • a small classifier learns your categories from those representations;
  • it trained in under a minute in these runs, without a dedicated training GPU.
• Dedicated AI Image Model (Business with Vision):
  • the AI itself is trained on your photos, on a GPU;
  • it took 33 to 54 minutes to train here, including a few minutes waiting for a GPU, and it is built for differences a ready-made model may not separate well.

Both approaches below are options in StayCharted AI Model Trainer. Damage accuracy in the main table measures all six categories; the separate damaged-or-not figure is a binary aggregation of those answers.

* Balanced accuracy is mean recall across categories, so a model that does well only on the common categories doesn't score well on it.

The default review cutoff is 0.5 for the AI Image Classifier Model and 0.9 for the Dedicated AI Image Model. Below it, an answer is marked for review. Projected after review assumes a person corrects every flagged mistake without introducing errors. Human review was not measured in this test.

Damage: the questions behind the six categories

• Front or rear was never confused. In this dataset, front and rear photos are easy to tell apart, so this says little about harder photos.
• Breakage versus crushed is where most mistakes are, and where people disagree most too. That's what the review step is for.

Part: where the mistakes are

• Front wheel and back wheel are most of the mistakes: a cut-out wheel looks the same either way, without the car around it. The Dedicated model found only 52.5% of front wheels. Treat them as one part, "wheel", and both models do much better (93.7% overall for the Dedicated model).
• Every other part was 80% or better for the Dedicated model. License plates, mirrors, tail lights and rocker panels were 97.5% or better.

Damage type: common versus rare

• Dent against scratch and dent against broken part are the most common mix-ups.
• The rare types were mostly called scratch. With 41 to 67 training pictures each, against about 300 for the common types, the models didn't learn them. Add representative examples, retrain and check whether rare-category recall improves.

Each answer comes with a confidence, and AMT sends the less confident ones to a person in the Review Queue. The customer picks the cutoff. Dedicated AI Image Model:

• Damage and part: checking about 1 picture in 4 or 5 projects 96–97% accuracy, assuming every flagged error is corrected.
• Damage type: the model is often unsure, and rightly so. At the default, half go to a person, and the projected result is 93.5%. Retraining on corrections may improve the next version; measure that change before publishing.

An agent that handles car photos needs two things: a model that knows your categories, and a person for the uncertain ones. For example:

• Insurers, first notice of loss: sort incoming claim photos into damaged or not, and front or rear. The confident ones route straight to the right queue; the uncertain ones go to an adjuster. Here, 77% were above the cutoff, with 98.4% damaged-or-not accuracy in that subset.
• Dealer service and body shops, intake: tag which part a customer's photo shows, and what the damage looks like, before an advisor sees it. Part recognition is strong. Damage type needs a person on about half today, and needs more representative examples and evaluation.
• Vehicle makers, warranty and quality: group returned-part or field photos by part and damage type, then look at what's growing. Start with the part model; add damage types as examples come in.

Start with the AI Image Classifier Model, which trained in under a minute in these tests, and see how it does on your own photos. Move to the Dedicated AI Image Model where it matters: here it caught 20 more damaged cars than the classifier did, and was more accurate on all three tests.

Judge it on the mistakes that matter to you. A damaged car passed as undamaged costs more than the reverse. AMT's validation report shows accuracy by category so you can check exactly that, and the review cutoff decides how many a person sees.

• One run per model, on public datasets. Your photos, lighting, angles and categories will give different numbers. Validate on your own held-out photos before relying on a model.
• The photos' original source isn't stated by either dataset. Both are published under open licences (MIT and CC0). We show no photos from them here.
• Front or rear at 100% reflects this dataset, not a hard problem solved.
• Cut-outs aren't customer photos. A customer's close-up of a dent shows more around it than a tight cut-out, which may make damage type easier in practice. It may also include more than one damage type.
• Rare damage types had too few examples to judge them one by one: 11 to 17 test pictures each. They're reported together.
• Speed isn't compared here. Training times include time waiting for a GPU, and fill times depend on how busy the service was.

• Run through StayCharted AI Model Trainer (amt.staycharted.com) on 21 October 2026, in a dedicated benchmark account on the Business plan with Vision, exactly as a customer would:
  1. upload a ZIP of sorted photos, one folder per category;
  2. review data quality;
  3. train;
  4. publish;
  5. fill a ZIP of test photos;
  6. score the filled answers against the held-out labels.
• AI Image Classifier Model: SigLIP 2 image embeddings plus a linear classifier.
• Dedicated AI Image Model: Qwen3.5-4B fine-tuned on the training photos, on a SageMaker GPU.

Compare fruit grading and cashew defects, or see how the image classifier uses SigLIP 2.

• DrBimmer, comprehensive-car-damage. MIT license.
• Humans in the Loop, Car Parts and Car Damages. CC0 1.0; polygon annotations by Beetroot Academy trainees.
• SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features, Tschannen et al., 2025.

Can AI classify car damage from photos?

In this test, StayCharted’s Dedicated AI Image Model achieved 94.0% damaged-or-not accuracy on 552 held-out public car photos. Across the six original damage categories it scored 86.4%. Validate on your own photos and review uncertain answers.

How much human review did the car damage model need?

At the default 0.9 cutoff, 23.0% were flagged. The remaining 77% had 98.4% damaged-or-not accuracy; four damaged cars were accepted as undamaged. Correcting all flagged errors would yield a projected 98.7% binary accuracy, not a measured reviewer result.

Can it recognize vehicle parts and rare damage types?

The Dedicated AI Image Model scored 89.9% across 21 cropped exterior parts and 70.2% across eight damage types. The four rare damage types scored only 19.0% collectively on 58 test crops, showing the need for representative examples.