Microsoft Common Objects in Context with 20% of Noisy Correspondence and 1K test data
This dataset is based on MS COCO that have 20% of data randomly shuffled to simulate noisy correspondence.