THE CHALLENGE
Our client is a global leader in developing deep learning software for Autonomous Driving (AD) and Advanced Driver Assistance Systems (ADAS). Their core challenge was managing the explosive growth and hyper-accuracy demands of the training data pipeline. To safely train Level 4 and Level 5 autonomous models, every single data point must be accurate to the sub-centimeter level.
Key challenges included:
- Precision: ADAS models rely on complex data types, including 3D Points, Bounding Boxes, Semantic Segmentation, and Bird’s Eye View (BEV) labelling. Errors in this data directly translate to perception failures and safety risks in production vehicles.
- Complexity of Sensor Fusion: Managing and accurately labeling vast datasets captured by multiple sensors (RGB cameras, LiDAR point clouds, radar) required annotators with specialized spatial reasoning skills far beyond typical 2D image labeling.
- Massive Volume and Velocity: To maintain their competitive edge, the client required annotation volume to scale significantly. The BPO partner had to guarantee this exponential growth while sustaining an accuracy rate of 96% or higher.
Our solutions
Expertise in Complex Labeling Types
Our specialized annotators mastered the full spectrum of required services: 3D Point Cloud and Bounding Box, Semantic Segmentation, and Bird’s Eye View.
Multi-Layered Quality Control (QC/QA)
To exceed the 97% accuracy requirement, every dataset went through multiple verification steps, including Consensus Scoring and Tool Agnostic Expertise.
THE RESULTS
Annotation Throughput Scale
Increased by 400%
Project Accuracy Rate
Reached 96% on average after 2 months
Data backlog
Reduced by 70% after 2 months
