KVQA: Knowledge-aware Visual Question Answering

KVQA is a visual question-answering dataset created by Sanket Shah, Anand Mishra, Naganand Yadati, and Partha Pratim Talukdar at IIIT Hyderabad / IISc. It is designed for questions about named entities in images that require world knowledge and reasoning over a knowledge graph.

The original release contains 24,602 images, approximately 183,000 manually verified question-answer pairs, more than 18,000 named entities, and a reference-image collection for visual named-entity linking. It also includes five image-level train/validation/test split assignments, paraphrased questions, question types, Wikidata QIDs, and preprocessed knowledge-graph facts.

Dataset contents

Path	Description
data/dataset.json	Image metadata, captions, named entities, Wikidata QIDs, questions, paraphrases, answers, question types, and split assignments
data/KVQAimgs.tar.gz	Main dataset images
data/KVQArefImgs.tar.gz	Reference images associated with Wikidata QIDs
data/KVQAkgFacts.tar.gz	Preprocessed closed-world KG facts and QID-to-name mapping
data/updates.txt	Upstream release notes

The JSON object is keyed by image ID. Each image record includes imgPath, Qids, NamedEntities, wikiCap, Questions, ParaQuestions, Answers, Type of Question, and split.

Intended use

This dataset is intended for research and educational work on visual question answering, visual named-entity linking, knowledge-graph reasoning, and multimodal learning. Users should review the source and rights status of the individual media files before using the dataset in any public, commercial, or redistributed application.

Maintained mirror and research extensions

This mirror is maintained by Wael AbdAlmageed to preserve access to KVQA and help the community continue research on knowledge-aware visual question answering. The published NeSy 2026 paper SoftReason: A Fully Differentiable Deductive Reasoning Architecture over High-Dimensional Perceptual Data instantiates a differentiable reasoning architecture on KVQA.

Researchers are encouraged to download the data, reproduce published results, report annotation or packaging issues, and contribute improved loaders, benchmarks, baselines, and reasoning methods. If you build on this mirror or use KVQA in work related to SoftReason, please cite both the original KVQA paper and the SoftReason paper.

For discussions, corrections, and proposed improvements, use the Hugging Face dataset discussions.

Provenance and licensing

This is a community-prepared mirror of the original KVQA 1.0 release. The authoritative project page is https://malllabiisc.github.io/resources/kvqa/. The upstream README states that the dataset images were harvested from Wikipedia, and the reference images were collected from Wikidata.

No explicit dataset license grant was identified on the original project page or upstream README. Accordingly, this repository does not relicense the dataset and does not grant rights beyond those held by the respective upstream rightsholders. The Hugging Face metadata uses license: other to avoid implying an open license.

In particular:

• Image rights can differ by file and may be held by third parties; Wikipedia hosting does not by itself establish a single license for every image.
• Wikidata data and reference images may have separate terms and source-specific restrictions.
• Dataset annotations and preprocessing are attributed to the KVQA authors; no additional license is asserted here.
• Before redistributing or using the media commercially, obtain permission where required and retain the applicable attribution and license information.

If the original authors provide a formal license or updated redistribution terms, this card should be amended to reflect them. For rights questions, the upstream README lists anandmishra@iisc.ac.in as a project contact.

Citation

Please cite the original dataset paper when using KVQA:

@InProceedings{shahMYP19,
  author    = "Sanket Shah, Anand Mishra, Naganand Yadati and Partha Pratim Talukdar",
  title     = "KVQA: Knowledge-Aware Visual Question Answering",
  booktitle = "AAAI",
  year      = "2019",
}

If you use this maintained mirror, reproduce the SoftReason experiments, or extend this line of research, please also cite:

@inproceedings{abdalmageed2026softreason,
  author    = {AbdAlmageed, Wael},
  title     = {SoftReason: A Fully Differentiable Deductive Reasoning Architecture over High-Dimensional Perceptual Data},
  booktitle = {Proceedings of the 20th Conference on Neurosymbolic Learning and Reasoning},
  year      = {2026},
  publisher = {Proceedings of Machine Learning Research},
  volume    = {284},
  url       = {https://arxiv.org/abs/2607.20402},
}

Sources

• Original project page: https://malllabiisc.github.io/resources/kvqa/
• Original README and download documentation: https://malllabiisc.github.io/resources/kvqa/files/README.txt
• Original release updates: https://malllabiisc.github.io/resources/kvqa/files/updates.txt
• Paper: https://ojs.aaai.org/index.php/AAAI/article/view/4915

Disclaimer

The upstream release notes that the dataset is human verified but may contain minor human errors. This mirror is provided for preservation and research discovery; users are responsible for complying with applicable law, source terms, and rights-holder requirements.