Machine Learning on Images with Noisy Human-Centric Labels
When humans describe images, they have a reporting bias, in that the report only what they consider important. Thus, in addition to considering whether something is present in an image, one should consider whether it is also relevant to the image
before labeling