To accomplish this, we use deep-learning-enhanced automation, enabling us to annotate data in a faster, more reliable way. We begin with a team of human annotators who do the first iteration — identifying objects like trees, cars, pedestrians, and bicyclists seen on the roads by our vehicles and sensor kits — and couple this with various engineering optimizations. We prioritize UI and workflow efficiency for our annotation team, building in features such as easily “scrubbing” backwards or forwards through time by dragging the cursor, and automatic interpolation of data between frames. Efficiencies such as these improve the rate and accuracy of the data annotation process.
And if it works, you've created successful autonomous vehicles. And if it doesn't, you've created 400 hours of busywork for every run to 7-11.