The autonomous driving industry is innovating at a rapid pace. From integrating advanced driver-assistance and monitoring systems to providing full-scale automated driving experiences, high-quality datasets form the cornerstones of these complex systems. TELUS Digital’s AI Data Solutions help automotive machine learning (ML) teams accelerate computer vision innovations at scale.
The challenge
Our client, one of the largest European automobile manufacturers, builds cutting-edge L4 & L5 autonomous driving technologies for improved mobility systems. The client’s perception team builds complex ML models to detect objects in large image sequences collected via multi-camera sensors. These safety-critical models require highly-accurate object detection datasets with 2D bounding box and 3D cuboid annotations. They needed a solution to optimize their pipeline with a continuous supply of high-quality datasets that could train their perception stacks quickly and effectively with minimal supervision of data labeling workflows.
The TELUS Digital solution
As part of the project scope, our client needed a data annotation platform that supported different label formats for multi-sensor data. Ground Truth Studios, TELUS Digital’s multi-faceted data annotation platform, helped the client eliminate the hassle of dealing with multiple labeling partners to quickly create large volumes of diverse datasets required for faster model training and enhancements.
TELUS Digital delivered:
- Faster and more accurate labeling for multi-sensor data via ML-assisted annotation tools
- Seamless data integration with robust APIs and Python SDKs
- High-security standards with a SOC-2 certified platform
- GDPR compliance via our proprietary image anonymizer that automatically blurs out personal identifiers from images
- 99% average recall rates for all outputs via specialized automation and quality control features that substituted for repetitive, high-effort tasks
- Robust training that enabled over 500+ annotators to execute complex labeling tasks
- Real-time annotator productivity tracking including instant feedback mechanisms
- Custom infrastructure engineering to support unique edge cases
- High quality and affordable data at scale using efficient and streamlined workflows
- Highly flexible ad hoc project support to accommodate changing requirements
The results
TELUS Digital’s solution quadrupled the client’s productivity in 60 days and generated over 3 million labels per month, while maintaining a 99% average recall for all outputs. As a result, their perception team improved their model accuracies and implemented faster experimentation for better outputs.