SpookyBench-v2

Dataset Description

SpookyBench-v2 is the updated version of spooky-bench benchmark, designed to evaluate vision-language models' capabilities for temporal understanding. The dataset contains both static and dynamic videos with various noise densities and speckle sizes, specifically designed to test models' temporal perception abilities. Check our paper for more details.

Key Features

• Human accuracy around 98% while all of the models including closed source one gets 0% accuracy.
• Noise Robustness Testing: Videos with varying noise densities (10-90) and speckle sizes
• Temporal Understanding: Mix of static and dynamic content to test temporal perception
• Diverse Categories: Images, words, shapes, real-world videos, and dual-object scenarios
• Movement Type Classification: Explicit labeling of static vs. dynamic content

Dataset Statistics

• Total Videos: 1,127
• Total Frames: 451,649
• Static Frame Videos: 394
• Dynamic Videos: 733

Categories Breakdown

• dual_images: 108 videos
• images: 404 videos
• shapes: 28 videos
• videos: 57 videos
• words: 530 videos

Dataset Structure

  SpookyBenchDatasets/ 
  ├── metadata.csv # Dataset metadata with all annotations 
  ├── images/ # Static single-object videos 
  ├── words/ # Static word videos 
  ├── shapes/ # Static geometric shape videos 
  ├── videos/ # Dynamic real-world depth-map-based videos 
  └── dynamic/ # Dynamic synthetic motion videos 
      ├── images/ # Single object with dynamic motion 
      ├── words/ # Words with dynamic motion 
      └── dual_images/ # Two objects with dynamic motion

Metadata Schema

The metadata.csv file contains the following columns:

Column	Description
file_name / Path	Relative path to video file
Category	Video category (images/words/shapes/videos/dual_images)
Label	Object/word labels in the video (list format)
Noise Density	Noise level from 10 to 90
Speckle Size	Speckle size parameter (1, 3, etc.)
Frame Count	Total number of frames
FPS	Frames per second
Duration	Video duration in seconds
Movement Type	static or dynamic

Usage

Loading the Dataset

  from datasets import load_dataset
  
  # Load the entire dataset
  dataset = load_dataset("mukul54/spookybench-v2")
  
  # Access a video
  example = dataset["train"][0]
  video_path = example["file_name"]
  label = example["Label"]
  noise_level = example["Noise Density"]
  movement = example["Movement Type"]
  
  print(f"Video: {video_path}")
  print(f"Label: {label}")
  print(f"Noise: {noise_level}, Movement: {movement}")
  Filtering by Category
  # Filter static images only
  static_images = dataset.filter(
      lambda x: x["Category"] == "images" and x["Movement Type"] == "static"
  )
  
  # Filter dynamic videos only
  dynamic_videos = dataset.filter(
      lambda x: x["Movement Type"] == "dynamic"
  )
  
  # Filter high noise videos
  high_noise = dataset.filter(
      lambda x: x["Noise Density"] >= 70
  )

Example Use Cases

• Noise Robustness Evaluation: Test model performance across different noise levels
• Temporal Understanding: Compare model accuracy on static vs dynamic content
• Zero-Shot Classification: Evaluate on diverse object categories in extreme cases
• Security and Captcha Verification: This can provide an ideal captcha for websites.

Citation

If you use this dataset in your research, please cite us using bibtex:

  @article{upadhyay2025time,
    title={Time Blindness: Why Video-Language Models Can't See What Humans Can?},
    author={Upadhyay, Ujjwal and Ranjan, Mukul and Shen, Zhiqiang and Elhoseiny, Mohamed},
    journal={arXiv preprint arXiv:2505.24867},
    year={2025}
  }

Contact

For questions or issues regarding the dataset, please open an issue on the dataset repository or contact the authors.