The main thing I took away from this survey is that the state of the datasets researchers pass around is kind of terrible. None of them are very large. The biggest one looks to be a pile of manga of dubious copyright status.<p>I find myself wondering if it might be worth spending some resources reaching out to independent comics creators to get copies of their source files, which will often have consistent methods of organization and labeling across hundreds of pages, as well as machine-readable text, and building a new dataset or expanding existing ones. Some time parsing the propriety formats these are mostly going to be in (largely Photoshop, Illustrator, and Manga Studio) and/or pulling data out via the program’s scripting/plug-in interfaces might prove worthwhile.<p>Reach out to “collectives” like Hiveworks or Spiderforest, you can find a lot of artists working at a pro level that way.