vqacovers

Textbook Cover VQA

A Visual Question Answering Dataset for Structured Understanding of Textbook Covers

Textbook Cover VQA is a Visual Question Answering (VQA) dataset developed for research on structured understanding of text-rich academic textbook covers. The dataset is designed to evaluate vision–language models (VLMs) on textual understanding, academic-domain recognition, spatial grounding, and visual content understanding.

Dataset Details

Dataset Description

Textbook Cover VQA consists of 1,510 textbook-cover images covering 9 academic domains. Each textbook-cover image is associated with 8 predefined VQA questions, resulting in 12,080 question-answer annotations.

The dataset captures four major types of information:

• Textual: title, author, and publisher
• Domain: primary academic subject/domain
• Spatial: locations of title, author, and publisher
• Visual: presence of illustration, diagram, or graphic

Project Information

• Guide: Apsana Parveen, Asst. Professor
• Affiliation: Department of Electronics and Communication Engineering, Bapatla Engineering College, India
• Project Batch: A11
• Team Members:
• B. Vinay Kumar — L24AEC592
• A. Akash — Y23AEC404
• A. Pujitha — Y23AEC410
• B. Avinash — Y23AEC413
• Language: English
• License: CC-BY-4.0

Dataset Sources

• Repository: https://huggingface.co/datasets/BVinay06/vqacovers
• Paper: To be added
• Demo: Not available

Uses

Direct Use

• Zero-shot evaluation of vision–language models
• Domain-specific fine-tuning and adaptation of VLMs
• Visual Question Answering on textbook covers
• Text-rich image understanding
• Spatial grounding of cover elements
• Academic subject/domain recognition
• Multimodal instruction-following evaluation
• Educational VLM benchmarking

Out-of-Scope Use

• General-purpose OCR benchmarking
• General natural-scene understanding
• Tasks unrelated to textbook-cover understanding
• Safety-critical or high-stakes decision making

Dataset Structure

The dataset contains 1,510 textbook-cover images and 8 questions per image.

The eight questions are:

1. What is the title of the textbook?
2. Who are the author(s) of the textbook?
3. What is the name of the publisher?
4. What is the primary subject or domain of this book?
5. Where is the title located on the cover page?
6. Where is the author's name located on the cover page?
7. Where is the publisher name located on the cover page?
8. Does the cover contain any illustration, diagram, or graphic?

Dataset Statistics

Each sample contains

• image — textbook-cover image
• image_id — unique image identifier
• Question fields — eight VQA questions
• Answer values — ground-truth annotations

Dataset Creation

Curation Rationale

The dataset was prepared to study how well vision–language models understand structured information present on academic textbook covers.

A single textbook cover can contain multiple information types, including title, author, publisher, academic domain, spatial placement, and graphics. The dataset therefore provides a controlled benchmark for multimodal understanding of text-rich educational images.

Data Collection and Processing

Textbook-cover images were prepared and organized into a structured VQA dataset. Annotation work was divided among the project team and organized using a shared spreadsheet.

The annotation process follows:

Image → Annotation → Review → Structured VQA Dataset

Common annotation categories and predefined spatial location options were used to maintain consistency across the dataset.

Annotations

Annotation Process

Each textbook-cover image is associated with eight fixed question-answer tasks covering:

• Title
• Author
• Publisher
• Subject / Domain
• Title Location
• Author Location
• Publisher Location
• Graphics

Answers were reviewed for consistency with the corresponding textbook cover.

Who are the Annotators?

The annotation work was prepared by the project team:

• B. Vinay Kumar
• A. Akash
• A. Pujitha
• B. Avinash

under the guidance of Apsana Parveen, Asst. Professor.

Personal and Sensitive Information

The dataset is intended to contain textbook-cover information and is not designed to contain personal, sensitive, or private information.

Intended Research Evaluation

The dataset is intended for evaluating vision–language models under a common textbook-cover VQA protocol.

The planned workflow is:

Textbook-Cover VQA Dataset → Zero-Shot VLM Evaluation → Domain Fine-Tuning → Fine-Tuned Evaluation → Performance Comparison → Error Analysis → Conclusion and Future Scope

Evaluation can include overall VQA accuracy, exact match for textual answers, task-wise accuracy, and macro-F1.

Bias, Risks, and Limitations

• The dataset is limited to textbook-cover images.
• It represents academic and educational material rather than general real-world scenes.
• Textbook layouts, fonts, visual styles, and image quality may vary.
• Some cover information may be partially visible or absent.
• Spatial annotations depend on predefined positional categories.
• The dataset should not be treated as a general-purpose VQA benchmark.

Recommendations

Users should treat this dataset as a benchmark for structured visual understanding of textbook covers and domain-adaptive VLM research.

For fair comparison, users should keep the question definitions and evaluation protocol consistent across models.

Citation

If you use this dataset in your research, please cite:

@dataset{textbook_cover_vqa_2026,
  title        = {Textbook Cover VQA: A Visual Question Answering Dataset for Structured Understanding of Textbook Covers},
  author       = {
    Vinay Kumar, B.
    and Akash, A.
    and Pujitha, A.
    and Avinash, B.
  },
  affiliation  = {Department of Electronics and Communication Engineering, Bapatla Engineering College, India},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/BVinay06/vqacovers}
}

Acknowledgement

We acknowledge the support of the Department of Electronics and Communication Engineering, Bapatla Engineering College, and the guidance of Apsana Parveen, Asst. Professor.