Autonomous Driving Auto-Labeling Pipeline
Design an end-to-end pipeline that turns driving logs into versioned machine-generated labels with human review and lineage.
Problem statement
For this exercise, assume immutable multi-sensor driving clips arrive in object storage. A workflow synchronizes sensors, runs several GPU perception models, fuses their predictions, applies quality rules, routes uncertain clips to human review, and publishes versioned labels for training.
The same clip may be reprocessed with a new model or calibration. Exclude model training and the annotation tool's pixel-level UI.
What to build
- Create a reproducible labeling run from clip, calibration, model, and policy versions.
- Execute synchronization, inference, fusion, quality checks, and optional review with retry-safe fan-out.
- Publish an immutable label version with complete lineage and quality evidence.
At scale
- Process 50 million one-minute clips per month with heterogeneous GPU requirements.
- Make accepted work durable within 500 ms and expose clip-level progress.
- Never let a partial run replace a previously published label version.
- Support targeted reprocessing by model, calibration, geography, and quality-rule version.