DEArt: Dataset of European Art

Large datasets that were made publicly available to the research community over last 20 years have been a key enabling factor for advances in deep learning algorithms NLP or computer vision. These are generally pairs of aligned image/manually annotated metadata, where images photographs everyday life. Scholarly and historical content, on other hand, treat subjects not necessarily popular general audience, they may always contain large number data points, new be difficult impossible collect. Some exceptions do exist, instance, scientific health data, but this is case cultural heritage (CH). The poor performance best models vision - when tested artworks coupled with lack extensively CH, fact artwork depict objects actions captured by photographs, indicate CH-specific dataset would highly valuable community. We propose DEArt, at point primarily an object detection pose classification meant reference paintings between XIIth XVIIIth centuries. It contains more than 15000 images, about 80% non-iconic, manual annotations bounding boxes identifying all instances 69 classes as well 12 possible poses human-like objects. Of these, 50 thus appear datasets; these reflect imaginary beings, symbolic entities categories related art. Additionally, existing include annotations. Our results show detectors domain can achieve level precision comparable state-of-art generic via transfer learning.