The "question_answer_reason.json" is the generated QA samples. It contains a list of question-answer samples. Each sample has the following field:
- "question": The question raw text.
- "answer": The answer raw text.
- "level": The inference step of this question.
- "KB": 0 or 1 that indicates whether this question is generated with extern knowledge base.
- "qtype": The question type described in the paper.
- "reason": A list that contains the used scene graph triplets from Visual Genome or knowledge triplets from FVQA("all_fact_triples_release.json").
- "image_id": The image id from Visual Genome.
- "question_id": The ID question of this question.
The "splits.json" contains questions' ID for our train/val/test split. It has the following keys:
- "train": a list of training "question_id"
- "val": a list of validation "question_id"
- "test": a list of test "question_id"
The images and their scene graph annoation can be downloaded from Visual Genome official website.
The extern knowledge based is provided by FVQA, and can be downloaded from dropbox. We use the "new_dataset_release/all_fact_triples_release.json" as the complete extern knowledge base.
 P. Wang, Q. Wu, et al. "FVQA: Fact-based visual question answering", IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017