Datasets:

howard-hou

/

OCR-VQA

like 24

Hi, thanks for sharing this dataset.

While i've checked the dataset. I found some missing_items.
'B001GNBU18', '823924335', '395291305', '1565842715', '399173013', '471927074', '1259589730', '143120581', '3882265868', '980011116', '1455773506', '471470538', '812969472', '73398144', '1492616737', '679754814', '007149572X', '089886772X', '1609785908', '1412716071', '865424306', '435302485', '1597752134'

And I found TextVQA/Cap-Style OCR tokens, which is useful but not provided in the meta OCR-VQA. Can you sharing more details about the ocr process?

The items were check with this script.

import json
import datasets
from collections import defaultdict
from tqdm import tqdm


def compare_qa_list(q_list_1, a_list_1, q_list_2, a_list_2):
    if not len(q_list_1) == len(q_list_2) == len(a_list_1) == len(a_list_2):
        return False
    if set(q_list_1) != set(q_list_2):
        return False
    if set(a_list_1) != set(a_list_2):
        return False
    qa_1 = {q:a for q, a in zip(q_list_1, a_list_1)}
    qa_2 = {q:a for q, a in zip(q_list_2, a_list_2)}
    for q in qa_1.keys():
        if not qa_1[q] == qa_2[q]:
            return False
    return True


def main_check_hf_datasets():

    hf_trainset = datasets.load_dataset("howard-hou/OCR-VQA", split="train")
    hf_valset = datasets.load_dataset("howard-hou/OCR-VQA", split="validation")
    hf_testset = datasets.load_dataset("howard-hou/OCR-VQA", split="test")
    
    src_dataset = json.load(open("datasets/OCR-VQA/dataset.json", "r"))
    
    unexpected_items, error_items = {}, {}, defaultdict(list)
    # check
    hf_all_image_id = []
    for split, dataset in [("train", hf_trainset), ("val", hf_valset), ("test", hf_testset)]:
        for item_id, item in enumerate(tqdm(dataset)):
            image_id = item["image_id"]  # 'image', 'image_id', 'questions', 'answers', 'ocr_tokens', 'ocr_info', 'title', 'authorName', 'genre', 'image_width', 'image_height', 'image_url', 'set_name'
            hf_all_image_id.append(image_id)
            if image_id in src_dataset:
                src_item = src_dataset[image_id]  # 'imageURL', 'questions', 'answers', 'title', 'authorName', 'genre', 'split'
                
                # check
                if not item["image_url"] == src_item["imageURL"]:
                    error_items["image_url_mismatch"].append((split, item_id))
                if not compare_qa_list(item["questions"], item["answers"], src_item["questions"], src_item["answers"]):
                    error_items["qa_mismatch"].append((split, item_id))
                if not item["title"] == src_item["title"]:
                    error_items["title_mismatch"].append((split, item_id))
            else:
                unexpected_items[image_id] = item
    missing_items = {img_id:src_dataset[img_id] for img_id in tqdm(set(src_dataset.keys()) - set(hf_all_image_id))}


if __name__ == "__main__":
    main_check_hf_datasets()