12K+
annotations delivered
95%+
accuracy sustained
90 days
turnaround
120
annotators deployed
1
The Challenge
Uber required frame-level video annotation across four simultaneous modalities 2D bounding boxes, 3D cuboids, polygon segmentation, and LiDAR point cloud labeling to train perception models for damage detection and obstacle recognition. Each video demanded approximately 400 annotations at the frame level, with near-zero tolerance for error.
- Managing four concurrent annotation types (2D, 3D, polygon, LiDAR) per video frame
- Maintaining frame-level consistency across long-form video sequences
- Achieving 99%+ accuracy with high-volume daily throughput (3–4 videos/day)
- Resolving recurring tool-related issues without impacting delivery timelines
Client
Uber
Industry
Autonomous Driving
Timeline
Ongoing production pipeline
2
Our Approach
- Deployed a dedicated team of 87 annotators, each handling ~4 videos with specialized modality training
- Designed a multi-modal workflow integrating 2D bounding boxes, 3D cuboids, polygon segmentation, and LiDAR point cloud annotation in a unified pipeline
- Built frame-level consistency checks to ensure annotation coherence across video sequences
- Established rapid issue resolution protocols to handle tool-related blockers without delivery delays
Quality Assurance
- Every annotation was independently reviewed by two different annotators across all four modalities
- Disagreements were flagged and escalated to senior QA reviewers for resolution
- Cross-modal consistency checks ensured 2D, 3D, polygon, and LiDAR annotations aligned correctly
- Only double-verified, consensus-approved annotations were delivered to the client
3
The Results
- Sustained 99–100% annotation accuracy across all four modalities
- Maintained ~1 hour turnaround per video with 400+ annotations
- Successfully scaled to 3–4 videos per day without quality degradation
- Eliminated cross-modal inconsistencies through rigorous multi-layer QA
Quality Assurance
OUTCOME
Delivered a production-ready, multi-modal annotation pipeline that directly accelerated Uber's autonomous driving perception model training reducing retraining cycles and enabling faster deployment of safety-critical AI systems.