Vehicle Damage Assessment

When a vehicle is involved in a collision, an adjuster reviews photos of the damage to determine what needs repair. This assessment can take days, and different adjusters may evaluate the same damage differently. Inconsistency leads to disputes, rework, and slower claim resolution.

Datature Vi trains a model on photos of damaged vehicles. You label images with damage descriptions (type, location, severity), and the model learns to produce consistent assessments from new photos. Policyholders can upload photos from their phone, and the model generates a preliminary report in seconds rather than days.

The model does not replace an adjuster's final decision. It produces a structured first-pass assessment that speeds up the review process and reduces variation between assessors.

For an interactive overview of this application, visit the vehicle damage assessment use case on vi.datature.com.

Common applications

Choose your task type

Annotation examples

Deploy and test

import json
from vi.inference import ViModel

model = ViModel(
    run_id="your-run-id",
    secret_key=".your-secret-key.",
    organization_id="your-organization-id",
)

result, error = model(
    source="vehicle_photo.jpg",
    user_prompt="Assess the visible damage on this vehicle.",
    generation_config={"temperature": 0.0, "do_sample": False}
)

if error is None:
    report = json.loads(result.result)
    for item in report["damages"]:
        print(f"{item['panel']}: {item['damage_type']} ({item['severity']})")

Training tips

Photograph from standard angles: train on the same angles policyholders will use (front, rear, both sides, close-up of damage). Consistent framing improves assessment accuracy.

Include a range of damage severity: from hairline scratches to total-loss collisions. If the model only sees major collisions in training, it will struggle with minor damage.

Use your organization's damage taxonomy: align annotations with your claims system's categories (panel names, damage types, severity levels) so model output maps directly to your workflow.

Include weather and lighting variation: photos submitted by policyholders come in all conditions. Include sunny, overcast, and indoor parking garage lighting in your training data.

Next steps