metadata

license: cc-by-4.0
task_categories:
  - question-answering
language:
  - en
size_categories:
  - 10M<n<100M

Perceptual Constancy

Perceptual Constancy is a multimodal benchmark designed to evaluate high-level perceptual invariance in large vision-language models (VLMs). It probes a model’s understanding of physical and geometric stability under varying sensory appearances. This dataset is part of the Grow AI Like a Child benchmark initiative.

🧠 Dataset Overview

The Perceptual Constancy dataset focuses on appearance-invariant reasoning using both static images and short video clips. Each question tests whether the model can generalize consistent properties across transformations such as viewpoint, color, orientation, size, or occlusion.

The dataset contains:

253 samples
Two modalities: image or video
Two question formats: multiple-choice (MC) or true/false (TF)

📁 Dataset Format

Each sample includes:

Field	Description
`Index`	Unique ID (e.g., `a0031`)
`Data.Type`	Either `image` or `video`
`Qustion.Type`	Either `MC` or `TF`
`Sec..Label`	Integer from 1 to 3 (see section mapping below)
`Question`	Natural language question with embedded options (for MC)
`Correct.Answer`	The correct response (e.g., `A`, `B`, `Yes`, `No`)

🔢 `Sec..Label` Categories

Label	Category
1	Color Constancy
2	Size Constancy
3	Shape Constancy

📂 Folder Structure

data/
├── data.csv
├── images/
│   ├── *.png / *.jpg / *.avif
│   └── metadata.jsonl
├── videos/
│   ├── *.mp4 / *.gif / *.mov
│   └── metadata.jsonl

The metadata.jsonl files store structured entries for each modality.
.gif files are stored in videos/ and marked as media_type = video.

💡 Example

{
  "file_name": "a0033.JPG",
  "media_type": "image",
  "question_type": "TF",
  "sec_label": 1.0,
  "question": "In the picture, has the actual color of the bridge itself changed?",
  "correct_answer": "No"
}

📚 Citation

If you use this dataset, please cite:

@misc{sun2025probingperceptualconstancylarge,
      title={Probing Perceptual Constancy in Large Vision Language Models}, 
      author={Haoran Sun and Suyang Yu and Yijiang Li and Qingying Gao and Haiyun Lyu and Hokin Deng and Dezhi Luo},
      year={2025},
      eprint={2502.10273},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2502.10273}, 
}

🤝 Acknowledgments

This dataset is developed by the Grow AI Like a Child community to support structured.