The COCO metadata provides annotations for each image in two forms, masks and captions. Both of these methods of annotation were made by hired Amazon Mechanical Turks and are showcased here.
