When the researchers discovered their algorithm was having trouble labeling pedestrians, cars, and other obstacles, they went back to look at their input data and discovered that most of the errors could be traced back to labeling errors that humans had made in the training dataset.4
