Uber Crosses the road

A good example of the complexity of data mining / machine learning data.  Via O'Reilly.

Inside Uber ATG’s Data Mining Operation: Identifying Real Road Scenarios at Scale for Machine Learning    By Steffon Davis, Shouheng Yi, Andy Li, and Mallika Chawda

How did the pedestrian cross the road?

Contrary to popular belief, sometimes the answer isn’t as simple as “to get to the other side.” To bring safe, reliable self-driving vehicles (SDVs) to the streets at Uber Advanced Technologies Group (ATG), our machine learning teams must fully master this scenario by predicting a number of possible real world outcomes related to a pedestrian’s decision to cross the road. To understand how this scenario might play out, we need to measure a multitude of possible scenario variations from real pedestrian behavior. These measurements power performance improvement flywheels for:

Perception and Prediction: machine-learned models with comprehensive, diverse, and continuously curated training examples (improved precision/recall, decreased training time, decreased compute).
Motion Planning: capability development with scenario-based requirements (higher test pass-rate, lower intervention rate).

Labeling: targeted labeling jobs with comprehensive, diverse, and continually updated scenarios (improved label quality, accelerated label production speed, lowered production cost).
Virtual Simulation: tests aligned with real-world scenarios (higher test quality, more efficient test runs, lowered compute cost).

Safety and Systems Engineering: statistically significant specifications and capability requirements aligned with the real-world (improved development quality, accelerated development speed, lowered development cost).

With the goal of measuring a scenario in the real world, let’s head to the streets to study how pedestrians cross them. ... '