@Paul Robinson: Did you look at the number of mislabeled images in the udacity self driving car dataset? If you do not have an understanding of the subject yourself, you’re likely to make the same annotation errors which feed into a bad model.