Awesome

FVQA: Fact-based Visual Question Answering

It can be downloaded from here.

• ./Name_Lists: the txt files contain the train and test images' id in the dataset.
• ./new_dataset_release/Images: the image files used by this dataset, which are extracted from COCO and ImageNet datasets.
• ./new_dataset_release/all_fact_triples_release.json: the facts (triplets <e1, rel, e2>) corresponding to the visual concepts shown in the paper (object/scene/action), which are extracted from three knowledge bases, i.e., DBpedia, ConceptNet and Webchild.
  -- r: the unique id of entity rel (relationship)
  -- e1: the unique id of entity e1
  -- e2: the unique id of entity e2
  -- KB: the source of knowledge base
  -- e1_label: the textual string describing entity e1
  -- e2_label: the textual string describing entity e2
  -- surface: the textual description of the fact
• ./new_dataset_release/all_qs_dict_release.json: the collected questions and their associated supporting facts.
  -- fact_surface: the textual description of the supporting fact
  -- ans_source: the source of the answer (Image or KB)
  -- question: the collected question
  -- answer: the answer to the question
  -- img_file: the image filename corresponding to the question
  -- visual concept: the visual concept corresponding to the supporting fact
  -- fact: the unique id of the supporting fact
  -- question id: the unique id of the question