Autonomous vehicle sensor data annotation for a German automotive OEM

Client profile

The client is a Stuttgart-headquartered automotive OEM with global revenue of approximately €38B and a well-established autonomous driving research and commercialization program. The client operates one of the industry's largest test fleets, generating multi-petabyte-scale sensor data (camera, LiDAR, radar, ultrasonic) requiring precise annotation to train perception models. The client's Stuttgart-based AI team includes 340 engineers, but internal annotation capacity is intentionally limited in favor of external annotation partnerships.

The challenge

The client's L4 autonomous driving program required approximately 8 million professionally-annotated frames per quarter across camera bounding boxes, semantic segmentation, LiDAR point cloud labeling, and 3D cuboid labeling. At engagement start, the client was working with two US-based annotation vendors delivering approximately 3.4 million frames per quarter combined, at an average unit cost of €3.20 per annotated frame. Both vendors had recurring quality issues that required rework, effectively pushing the true cost per usable frame closer to €4.10.

Beyond volume, the client had specific technical requirements no vendor was meeting adequately: inter-annotator agreement above 98% for safety-critical labels (pedestrians, cyclists, traffic signs), consistency across day/night/weather conditions, and full traceability from raw sensor data through annotation to model training runs for regulatory audit purposes. Both existing vendors used ad-hoc quality processes.

The client had also encountered a scaling ceiling. Both existing vendors had communicated that expanding capacity beyond ~2M frames per quarter each would require 6-9 month lead times. The client's L4 program timeline required tripled annotation capacity within 12 months.

Why Corpshore Vietnam

Corpshore Vietnam won a significant portion of the client's annotation program on three qualifications: our documented ability to scale annotation teams by 4x within 6 months (demonstrated on prior engagements), our proprietary annotation quality methodology delivering 99.6% inter-annotator agreement on bounding boxes (verified during a 20,000-frame trial), and our unit pricing at €1.10 per annotated frame (approximately 66% below the incumbent vendor pricing).

The client's evaluation specifically noted our annotation quality traceability, where every annotated frame is linked to the specific annotator, timestamp, review status, and QA cycle, meeting the client's audit requirements without the client-side rework the incumbents required.

The engagement

128 annotators split between Hanoi (76) and HCMC (52), with 12 senior specialists handling complex 3D cuboid and LiDAR point cloud labeling, and 6 QA leads managing cross-team calibration. Team scaled from 20 annotators at month one to 128 by month eight. All annotators completed a 6-week paid training program covering automotive perception taxonomy, safety-critical label handling, and the client's specific annotation schema. Weekly calibration sessions with the client's Stuttgart AI team run 4:00 PM CET (10:00 PM ICT) on Fridays.

Approach and methodology

Progressive complexity workflow. New annotators start on 2D bounding boxes and progress to semantic segmentation, LiDAR point cloud, and 3D cuboid labeling as they demonstrate proficiency. Only specialists with 6+ months of demonstrated quality handle safety-critical labels (pedestrians, cyclists).

Dual-annotator + specialist adjudication. Every safety-critical label is dual-annotated. Disagreement above the client's quality threshold triggers senior specialist review. This drove 99.6% inter-annotator agreement on bounding boxes and 98.9% on segmentation, both exceeding the client's requirement.

Full traceability infrastructure. Every annotated frame is metadata-tagged with annotator ID, timestamp, review cycle, and confidence score. The metadata is exported to the client's model training platform, enabling audit-grade traceability from raw sensor input to model behavior.

Results

By month 8, quarterly throughput reached 4.2 million frames (26% above the client's target), with quality metrics exceeding all thresholds. Total annotation program cost dropped by an estimated 62% versus the incumbent vendor baseline. The client's L4 program is on track for its 2027 commercialization milestone with the client explicitly citing annotation partner quality as a contributing factor.

The engagement has expanded to include Corpshore AI's synthetic data generation services for edge-case scenarios where real-world data is scarce (extreme weather, rare pedestrian behaviors, unusual construction zones).

Enduring value

The annotation traceability infrastructure and quality methodology developed for this engagement have influenced the client's internal AI data governance framework. The framework is now being applied to non-autonomous vehicle AI programs across the client's software organization.