Oztobuzz/tiny-gqa-bounding-box-detailed

Tiny GQA Bounding Box (Detailed)

Dataset Summary

Tiny GQA Bounding Box (Detailed) is a subset of the GQA dataset enriched with object-level grounding information. Each example pairs a question-answer instance with a key object, including its bounding box and semantic label, derived from scene graphs.

This dataset is designed for research in:

• Visual Question Answering (VQA)
• Multimodal reasoning
• Grounded reasoning and interpretability
• Object-centric evaluation of vision-language models

---

Dataset Structure

Each example contains:

• question_id: Unique identifier for the question
• image_id: Identifier linking to the image
• question: Natural language question
• answer: Ground truth answer
• key_object_id: Object ID from the scene graph associated with the answer
• key_object_name: Semantic label of the key object
• object_bounding_box:
  • x, y: Top-left coordinates
  • w, h: Width and height
• question_types:
  • structural: Question structure category
  • semantic: Semantic category
  • detailed: Fine-grained type
• image: The corresponding image

---

Data Source and Construction

This dataset is constructed from:

• GQA question annotations (val_balanced_questions.json)
• GQA scene graphs (val_sceneGraphs.json)
• Image data from a Hugging Face-hosted GQA dataset

Processing Steps

1. Iterate over GQA validation questions
2. Extract the key object from annotations.fullAnswer
3. Retrieve:
   • Object metadata from scene graphs
   • Corresponding image from the dataset
4. Construct a structured example with:
   • Question-answer pair
   • Grounded object (name + bounding box)
5. Filter out samples where fullAnswer is missing
6. Limit dataset size (~18k samples)

---

Dataset Size

Split	Examples
Train	18,744

---

Intended Use

This dataset is particularly useful for:

• Evaluating whether models attend to the correct object when answering questions
• Studying grounding behavior in multimodal LLMs
• Training or benchmarking object-aware VQA systems
• Supporting research in Explainable AI (XAI) for vision-language models

---

Limitations

• Only a subset of GQA is included (~20k samples before filtering)
• Each sample contains one key object, which may not capture full reasoning chains
• Reliance on fullAnswer annotations may introduce bias toward certain object types
• No guarantee that the selected object is the only relevant evidence

---

Licensing

License is currently unknown.
Users should refer to the original GQA dataset for licensing considerations.

---

Citation

If you use this dataset, please consider citing the original GQA paper:

@inproceedings{hudson2019gqa,
  title={Gqa: A new dataset for real-world visual reasoning and compositional question answering},
  author={Hudson, Drew A and Manning, Christopher D},
  booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
  pages={6700--6709},
  year={2019}
}