According to a recent study from MIT, 10 of the most commonly-used computer vision datasets have errors that are “numerous and widespread.”