jae-minkim/vstar-bench-with-bbox

vstar-bench with Bounding Box Annotations

Dataset Description

This dataset extends lmms-lab/vstar-bench by adding bounding box annotations for target objects. The bounding box information was extracted from craigwu/vstar_bench and mapped to the lmms-lab version.

Key Features

• Visual Spatial Reasoning: Tests understanding of spatial relationships in images
• Bounding Box Annotations: Each sample includes target object bounding boxes
• Multiple Choice QA: 4-way multiple choice questions about spatial relationships
• High Coverage: 100.0% of samples have bounding box annotations

Dataset Statistics

• Total Samples: 191
• Samples with Bbox: 191
• Coverage: 100.0%

Dataset Structure

Each sample contains:

• image: PIL Image object
• text: Question text (includes target object mention)
• bbox_target: Bounding box coordinates [x, y, width, height]
• target_object: Name of the target object
• label: Correct answer (A/B/C/D)
• question_id: Unique question identifier
• category: Question category (e.g., "relative_position")

Example

from datasets import load_dataset

dataset = load_dataset("jae-minkim/vstar-bench-with-bbox")
sample = dataset['test'][0]

print(f"Question: {sample['text']}")
print(f"Target Object: {sample['target_object']}")
print(f"Bounding Box: {sample['bbox_target']}")
print(f"Answer: {sample['label']}")

Usage

Load Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("jae-minkim/vstar-bench-with-bbox")

# Access test split
test_data = dataset['test']

# Iterate over samples
for sample in test_data:
    image = sample['image']
    bbox = sample['bbox_target']  # [x, y, width, height]
    question = sample['text']
    answer = sample['label']

Visualize Bounding Boxes

import matplotlib.pyplot as plt
import matplotlib.patches as patches

sample = dataset['test'][0]

fig, ax = plt.subplots(1, 1, figsize=(10, 8))
ax.imshow(sample['image'])

if sample['bbox_target']:
    x, y, w, h = sample['bbox_target']
    rect = patches.Rectangle(
        (x, y), w, h,
        linewidth=3, edgecolor='red', facecolor='none'
    )
    ax.add_patch(rect)
    ax.text(x, y-10, sample['target_object'], 
           color='red', fontsize=12, fontweight='bold')

ax.set_title(sample['text'])
ax.axis('off')
plt.show()

Applications

This dataset is particularly useful for:

1. Vision-Language Model Evaluation: Testing spatial reasoning capabilities
2. Attention Visualization: Analyzing if models attend to correct regions
3. Token Shifting Research: Redirecting high-norm tokens to target regions
4. Grounded QA: Question answering with explicit spatial grounding

Source Datasets

• lmms-lab/vstar-bench: Base dataset with questions and images
• craigwu/vstar_bench: Source of bounding box annotations

Citation

If you use this dataset, please cite the original vstar-bench paper:

@article{wu2023vstar,
  title={V*: Guided Visual Search as a Core Mechanism in Multimodal LLMs},
  author={Wu, Penghao and Xie, Saining},
  journal={arXiv preprint arXiv:2312.14135},
  year={2023}
}

License

Apache 2.0 (following the original vstar-bench license)

Acknowledgments

• Original dataset: lmms-lab/vstar-bench
• Bounding box source: craigwu/vstar_bench