Use Cases

Challenge

ML engineers training perception models for autonomous driving had no scalable way to annotate and select training data from the massive streams of camera and LiDAR sensor data generated during test drives. Data was imbalanced and unrepresentative, leading to gaps in model coverage and inconsistent performance across real-world scenarios.

Solution

• Scalable perception tagging pipeline processing camera and LiDAR data simultaneously across test fleets
• Time-windowed sampling produced balanced, diverse datasets for deep learning model training
• Spark APIs identified sensor data gaps, directly improving model accuracy and reliability

What we built

We designed and implemented a Perception Tagging Pipeline built on Apache Spark, creating a scalable object detection workflow that processed data across multiple camera feeds and LiDAR sensors simultaneously. Time-windowed sampling ensured diverse, balanced image selection, so ML engineers received representative datasets rather than redundant frames from similar conditions. We also built Spark APIs to analyze large-scale sensor data, surfacing training gaps and enabling targeted data collection to improve model accuracy and reliability.

Outcome

ML teams gained a reliable, automated pipeline for curating high-quality training data at scale. Model performance improved through better dataset balance, and lifecycle management became more structured, with clear visibility into where models were underperforming and what data was needed to close the gaps.