Using simple techniques, they found out that popular open source datasets like VOC or COCO contain up to 20% annotation errors in. By manually correcting those errors, they got an average error reduction of 5% for state-of-the-art computer vision models.