chandrabhuma/building_defect_vqa

Building Defect VQA Dataset

Dataset Description

This dataset is a Visual Question Answering (VQA) version of the original BD3 (Building Defect Dataset). It is designed for training and evaluating Vision–Language Models (VLMs) on building defect recognition tasks.

Each image is paired with a fixed question and a defect category as the answer.

---

Dataset Structure

Each sample contains:

• image: RGB image of a building surface
• question:

  "What type of building defect is visible in the image?"

• answer: One of the following defect classes:
  • algae
  • major_crack
  • minor_crack
  • peeling
  • plain
  • spalling
  • stain

---

Intended Use

This dataset is suitable for:

• Visual Question Answering (VQA)
• Vision–Language Model fine-tuning
• Prompt-based image classification
• Zero-shot / few-shot VLM evaluation

Compatible with models such as:

• LLaVA
• SmolVLM
• Pixtral
• Qwen-VL
• BLIP-style models

---

Source Dataset & Credits

This dataset is derived from the original BD3 (Building Defect Dataset).

Original Dataset

• Name: BD3 – Building Defect Dataset
• Authors: {Kottari, Praveen and Arjunan, Pandarasamy}
• Original Source: [https://github.com/Praveenkottari/BD3-Dataset]

Credit Statement

This VQA dataset is a reformatted version of the original BD3 dataset.
All image rights and original credit belong to the original authors.
This version only changes the data format to support VQA-style learning.

If you use this dataset, please cite the original BD3 dataset and paper.

---

License

This dataset follows the same license as the original BD3 dataset.
Please check the original source before using it for commercial purposes.

---

Citation

If you use this dataset, please cite:

@dataset{bd3_building_defect,
  title     = {BD3: Building Defect Dataset},
  author    = {Kottari, Praveen and Arjunan, Pandarasamy},
  year      = {2024},
  url       = {https://github.com/Praveenkottari/BD3-Dataset)}
}