GenAI – Enabling ADAS with High-Volume 3D Cuboid Annotations

SUMMARY

The project involved detecting and annotating 3D cuboid bounding boxes around moving objects—vehicles, pedestrians, and more—from ego vehicle camera feeds. Designed for ADAS (Advanced Driver-Assistance Systems), it aimed to enhance autonomous driving perception using high-volume, high-quality labeled data over 8 months.

THE CHALLENGE

• Managing annotation quality across 4M+ instances at scale.
• Maintaining 3D accuracy across varied object types and camera perspectives.
• Ensuring speed and consistency with 100K+ images per month.
• Aligning multi-object tracking from front and side angles.

SOLUTION

• Large-scale team: 75 annotators & 20 QC inspectors.
• Two-layered QA process: 100% QC in Layer 1, 30% sampling in Layer 2.
• Used client-provided platform tailored to cuboid annotations.
• 1-week incubation phase to align team skills and guidelines.
• Streamlined execution across 8 months for uninterrupted delivery.

KEY OUTCOMES

• Over 4 million accurate 3D cuboid annotations delivered.
• Sustained throughput of 100K+ annotated images per month.
• High-quality, ADAS-ready dataset to train and validate autonomous systems.
• Achieved consistent labeling accuracy through layered QA.