DocVQA – A Dataset for VQA on Document Images

Dataset Overview

• Name: DocVQA
• Task Type: Document Image Question‑Answering
• Source: Document images from the UCSF Industry Documents Library, covering printed, typed, and handwritten content.

Structure

• Total Questions: 50,000
• Total Images: 12,767
• Splits: Random 80‑10‑10 for training, validation, and test.
  • Training: 39,463 questions, 10,194 images
  • Validation: 5,349 questions, 1,286 images
  • Test: 5,188 questions, 1,287 images

Access

• The dataset can be downloaded from the “Downloads” tab of the RRC Challenge page.

Citation

@InProceedings{mathew2021docvqa, author = {Mathew, Minesh and Karatzas, Dimosthenis and Jawahar, CV}, title = {Docvqa: A dataset for vqa on document images}, booktitle = {Proceedings of the IEEE/CVF winter conference on applications of computer vision}, year = {2021}, pages = {2200--2209}, }