Data Annotation, Data‑Driven Development, and Decision‑Making in Autonomous Driving

The speaker, Pony.ai Tech Lead Song Hao, emphasizes that artificial intelligence requires large volumes of high‑quality data, and that fully leveraging this data is a core competitive advantage for AI companies.

Role of Data : Data drives development (providing training material and evaluation benchmarks), decision‑making (prioritizing safety, comfort, and operational issues), showcases company strength, and satisfies regulatory requirements.

Data Annotation : Key evaluation metrics include team size, efficiency vs. cost, and annotation quality. Additional important factors are task diversity (ability to handle lanes, traffic lights, obstacles, etc.) and flexibility for long‑tail scenarios.

Team Composition : A large, well‑structured team is essential, with clear roles for annotation, quality inspection, and re‑inspection to balance cost and efficiency.

Efficiency & Cost Control Techniques : ① Default obstacle size – using human‑machine interaction to set initial sizes. ② Automatic tracking extrapolation – the system predicts object positions across frames. ③ Automatic interpolation – the system fills in missing positions between annotated frames.

Non‑technical measures include salary incentives, organizational design for information flow, and balancing costs across annotation stages.

System Capability : The platform must remain stable under high‑frequency, large‑scale use, with robust testing environments, staged releases, and monitoring/incident response plans.

Data‑Driven Development : A feedback loop where machine intelligence improves annotation efficiency, which in turn enhances the models, creating a virtuous cycle. Key actions are distributed training/evaluation, handling annotation quality limits, and building a data index platform for balanced data selection.

Data‑Driven Decision‑Making : Analyzing road‑test events to identify problematic segments, modules, vehicles, and time windows, and presenting data in accurate, real‑time, user‑friendly visualizations tailored to different audiences (operations, executives, tech leads).

Tips for Improving Efficiency : Implement optimization solutions, measure annotator workflows, and iterate improvements based on observations.