ComTQA Dataset

1. Introduction

This dataset is a visual table question answering benchmark. The images are collected from FinTabNet and PubTables-1M. It totally includes 9070 QA pairs with 1591 images. The specific distribution of data is shown in the following table.

2. How to use it

• First，please download the FinTabNet and PubTables-1M from their original websites. The structure of collected data is formatted as follows,

  root
  └─FinTabNet
          ├─ pdf
          ├─ FinTabNet_1.0.0_cell_test.jsonl
          ├─ FinTabNet_1.0.0_cell_train.jsonl
          └─ ...
  └─PubTables-1M
          ├─ PubTables-1M-Structure
              ├─ images
              ├─ ...
          ├─ PubTables-1M-Detection
          ├─ ...
  

• Second, you can follow the steps below to extract the corresponding images.

  • For PubTables-1M, the key ''image_name'' in annotation.json represents the filename in the "./PubTables-1M/PubTables-1M-Structure/images".

  • For FinTabNet, the key "table_id" in annotation.json represents the same key in the file "FinTabNet_1.0.0_cell_test.jsonl". You could crop the table images from the original PDF with the annotations in "FinTabNet_1.0.0_cell_test.jsonl"

Citation

If you find this dataset useful for your research, please consider citing our work:

@inproceedings{zhao2024tabpedia,
 title={TabPedia: Towards Comprehensive Visual Table Understanding with Concept Synergy},
 author = {Weichao Zhao, Hao Feng, Qi Liu, Jingqun Tang, Binghong Wu, Lei Liao, Shu Wei, Yongjie Ye, Hao Liu, Wengang Zhou, Houqiang Li, Can Huang},
 booktitle = {Advances in Neural Information Processing Systems},
 year = {2024}
}