Vehicle Inspection AI Training Data – 1M+ Images Classified at Scale

LXT delivered consistent, granular damage severity labels across a high-volume image annotation program for a vehicle inspection AI platform — over one million images classified, fully integrated via API, and accepted without rejection.

The Challenge

A vehicle inspection AI platform specializing in automated damage assessment needed large-scale human-labeled image data to train and validate its models. The platform processed vehicle images sourced from automotive auctions and remarketing workflows, where accurate damage classification directly influenced downstream pricing and routing decisions. Training data of this kind requires not just volume, but precision: the difference between a vehicle with minor paint scuffs and one with a detached bumper is commercially significant, and ambiguous or inconsistent labels degrade model performance.

Specifically, the program had to:

• Classify each vehicle image against an 8-level damage severity scale, from no visible damage through to total loss, including intermediate states for partial panel damage at 25%, 50%, and 75% severity.
• Handle a high volume of images at sustained throughput, with results returned directly to the client’s systems via API.
• Correctly identify and flag invalid inputs: images that failed to load, were too low quality to assess, showed non-standard vehicle types such as motorcycles or commercial vehicles, depicted more than one vehicle, or where snow obscured the visible damage.
• Apply each damage rating independently per image, without inferring damage from other angles of the same vehicle.
• Maintain consistent label quality across a contributor pool large enough to sustain the program’s volume requirements.

Automated damage detection models at the time were not sufficiently reliable for the granularity the client required. Human judgment was necessary to distinguish between severity tiers — particularly at the intermediate thresholds — and to correctly handle edge cases that a taxonomy-driven classifier would misfire on.

The Solution

LXT built a structured image classification workflow around the client’s 8-level damage taxonomy, qualifying contributors through Gold Data testing and enforcing answer quality through majority vote validation.

Damage Taxonomy Built for Precision at the Threshold
The classification task used a carefully defined 8-level scale that separated cosmetic imperfections (paint chips, scuff marks, minor rust) from structural damage affecting individual panels, and structural damage from partial or total loss scenarios. Each level carried precise definitions and visual reference examples so contributors applied the same standard consistently across hundreds of thousands of vehicle inspection images. The taxonomy also covered non-classifiable inputs through a dedicated Image Error category and a Not Valid category for snow-obscured vehicles, wrong vehicle types, and multi-vehicle frames.

Qualified Contributors, Enforced by Gold Data
Contributors were qualified via Gold Data tests before accessing production tasks. Gold Data embedded pre-labeled reference images into the task stream and measured each contributor’s accuracy against ground truth. Those who did not meet the required standard were removed from the project, ensuring the active pool consisted only of contributors who had demonstrated correct application of the damage scale.

Majority Vote Quality Assurance
Each image was assessed by multiple contributors, with the final label determined by majority vote. Contributors were only paid for answers that aligned with the majority, creating a direct incentive for accurate rather than fast responses. This mechanism filtered out noise at scale without requiring manual review of every item.

API-Integrated Delivery
The program ran entirely via API. The client submitted images programmatically and received classified results directly into their systems, without manual file exchange. LXT absorbed the integration cost, and the resulting pipeline handled the full volume of the engagement without interruption.

Project Specifications

The exact numbers and rules behind the approach above, scannable at a glance.

• Task Type: Image classification — vehicle damage severity
• Damage Scale: 8 levels: Image Error · No Scratches or Very Minor Scratches · Some Small Dents or Dings · 25% Damaged · 50% Damaged · 75% Damaged · 100% Damaged · Not Valid
• Total Images Classified: 1,015,390
• Contributor Qualification: Gold Data tests; non-qualifying contributors removed from project
• Quality Mechanism: Majority vote; payment withheld for answers not matching majority
• Delivery Method: API integration; results returned directly to client systems
• Invalid Image Handling: Dedicated Image Error category (failed load, bad quality, non-vehicle content)
• Not Valid Criteria: Snow obscuring damage · non-standard vehicle types (motorcycles, commercial, industrial) · more than one vehicle in frame
• Per-Image Rule: Each image rated independently; no inference from other angles of the same vehicle
• Job Acceptance: 100% — all delivered jobs accepted by client

Contributor Task Flow

1. Pass Gold Data qualification test, demonstrating correct application of the 8-level damage scale before accessing production images.
2. View the vehicle image and confirm it loads correctly and shows a classifiable subject.
3. Select Image Error if the image fails to load, is too low quality, or does not show a vehicle.
4. Select Not Valid if the image shows snow obscuring the vehicle, a non-standard vehicle type, or more than one vehicle in focus.
5. For valid images, assess only the visible damage in that specific image without referencing other angles of the same vehicle.
6. Select the damage level that best matches the visible condition, using the defined scale and visual reference examples as the standard.

Outcomes

• Classified over one million vehicle images against an 8-level damage severity taxonomy, delivering a labeled dataset at a scale sufficient to train and validate automotive computer vision models for vehicle inspection AI.
• Achieved 100% job acceptance from the client, with no deliveries rejected across the full volume of the engagement.
• Ran the entire program via API, eliminating manual file transfer and integrating directly into the client’s data pipeline.
• Maintained label consistency through Gold Data qualification and majority vote validation, controlling quality at scale without a manual review bottleneck.
• The program concluded when the client’s vehicle inspection AI model reached sufficient maturity, with the LXT-labeled dataset serving as its training foundation.

Why LXT

• Contributor qualification via Gold Data testing — only contributors who demonstrated accuracy on the damage taxonomy accessed production work, ensuring label quality from the first image.
• Majority vote validation built into the task design, providing scalable quality control for vehicle inspection AI training data without manual review overhead.
• Full API integration: LXT absorbed the integration cost and delivered results directly into the client’s pipeline, with no manual file exchange at any point in the engagement.
• Sustained throughput at volume: capacity to classify over one million vehicle images with 100% client acceptance, demonstrating consistent output quality at scale.