High-quality vehicle annotation delivered ahead of schedule for a leading public safety technology company

LXT supplied 500k+ production-ready car and vehicle annotations across diverse real-world environments, completing the engagement one week early at 98%+ accuracy.

The Challenge

A leading public safety technology company needed high-quality vehicle annotation data to improve the performance of its AI models. The project covered cars and other vehicles captured in diverse, real-world environments, with particular emphasis on accuracy in challenging edge cases where standard car annotation pipelines typically underperform.

Specifically, the program had to:

• Deliver 2D bounding box vehicle annotations alongside structured attribute classification covering body type, make, visibility, occlusion, orientation, and license plate issuing state — covering passenger cars through to larger vehicle classes.

• Annotate aerial building images with geo coordinates, extending the scope beyond on-road scenarios.

• Sustain accuracy above 98% across a high volume of complex, edge-case imagery.

• Complete delivery on schedule without sacrificing quality at scale.

Off-the-shelf annotation tooling and generic crowd-sourced pipelines lacked the domain-specific qualification processes and adjudication workflows needed to sustain this accuracy level at volume. The client selected LXT on the basis of an existing, trusted relationship and a proven track record in high-quality AI data delivery.

The Solution

LXT designed a structured, two-pass vehicle / image annotation pipeline built around quality from the first task, not as a retrospective fix. The same system handled both car annotation tasks and broader vehicle categories within a single unified workflow.

Two-pass annotation with adjudication
Every annotation passed through a primary labelling stage followed by a dedicated adjudication pass to resolve discrepancies. The adjudication feature was launched pre-production, allowing the team to calibrate the workflow before full-scale execution began.

Annotator qualification and production gating
A rigorous qualification process ensured only top-performing annotators advanced to production work. This gating approach protected output quality and reduced rework downstream.

Targeted quality assurance
A stratified sampling QA process focused review effort on edge cases rather than applying uniform sampling across all tasks, making quality control both thorough and efficient.

Agile workflow management
The project team maintained an agile approach throughout, enabling real-time adjustments to workflows in response to evolving client requirements and contributing to cost efficiency across the engagement.

Project Specifications

• Annotation type: 2D bounding box vehicle annotation (cars and other vehicle classes)
• Attribute classification: Body type, make, visibility, occlusion, orientation, license plate issuing state
• Additional scope: Aerial building images with geo coordinates
• Total annotations: 500k individual annotations (across 110K images)
• QA method: Stratified sampling, edge-case prioritised
• Pipeline design: Two-pass: primary annotation + adjudication for discrepancies
• Annotator gating: Qualification process prior to production access
• Delivery outcome: Completed one week ahead of schedule

Outcomes

• Delivered 500k+ high-quality, production-ready annotations
• Completed the final production phase one week ahead of schedule at 98%+ accuracy.
• Launched an adjudication feature pre-production, validating the workflow before full-scale execution and reducing post-delivery revision cycles.
• Refined annotation guidelines and training materials in collaboration with the client, contributing to long-term quality improvements in their AI systems.
• Reduced costs through agile workflow adjustments while sustaining delivery commitments throughout the engagement.

Why LXT

• Established relationship: the client selected LXT on the strength of prior successful engagements, not a cold procurement process.
• Production-gated quality: annotator qualification and pre-production adjudication calibration prevented quality issues from reaching the delivery stage.
• Edge-case focus: stratified QA sampling concentrated review effort where accuracy risk was highest.
• Delivery discipline: agile project management enabled schedule and cost commitments to be met without compromising the 98%+ accuracy bar.