Perspective-Taking Dataset

Dataset Description

This dataset contains image-question pairs for perspective-taking tasks.

Dataset Statistics

• Training samples: 218
• Testing samples: 25
• Total samples: 243
• Concepts: concept_10_image, concept_10_multiimage

Dataset Structure

The dataset is organized into train and test splits for each concept. Each sample consists of:

• An image file (or multiple images)
• A question about the image
• An answer to the question (when available)

Data Fields

• id: Question identifier
• concept: The concept category the question belongs to
• question: The question text from question.txt
• answer: The answer text from answer.txt (when available)
• image: Filename of the primary image
• additional_images: Additional images if present
• additional_text: Additional text files with their content
• split: Train or test split

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("path/to/dataset")

# Access examples
sample = dataset["train"][0]
print(f"Question: {sample['question']}")
print(f"Answer: {sample.get('answer', 'No answer available')}")
print(f"Image path: {sample['image']}")

# If the sample has additional images
if 'additional_images' in sample and sample['additional_images']:
    print(f"Additional images: {sample['additional_images']}")

Citation

@article{gao2024vision,
  title={Vision Language Models See What You Want but not What You See},
  author={Gao, Qingying and Li, Yijiang and Lyu, Haiyun and Sun, Haoran and Luo, Dezhi and Deng, Hokin},
  journal={arXiv preprint arXiv:2410.00324},
  year={2024}
}

arxiv link: https://arxiv.org/abs/2410.00324