IconArt is a painting image dataset for weakly supervised object detection, containing 5,955 images from WikiCommons, 2,978 images in the training set, 2,977 images in the test set (for classification tasks), and 1,480 images in the test set are annotated with 7 categories of bounding boxes: "angle", "Child_Jesus", "crucifixion_of_Jesus", "Mary", "nudity", "ruins", "Saint_Sebastien".

The second version of the dataset has the same images as the first version, but with 10 categories of bounding box annotations: “angle”, “beard”, “capital”, “Child_Jesus”, “crucifixion_of_Jesusu”, “Mary”, “nudity”, “ruins”, “Saint_Sebastien”, “turban”.