Datasets:

mukul54
/

motion-text-classifier

license: mit
task_categories:
  - video-classification
  - zero-shot-classification
tags:
  - motion-detection
  - temporal-vision
  - speckle-noise
size_categories:
  - n<1K

Motion-Text Classifier Dataset

A binary classification dataset for distinguishing between pure motion and text-revealed-through-motion in speckle noise videos.

Dataset Description

This dataset contains 1000 videos (5 seconds each) designed to test motion-based perception:

Motion-only (500 videos): Pure background noise scrolling horizontally
Text videos (500 videos): Single words revealed through opposing foreground/background motion

All videos use identical noise parameters to ensure the only distinguishing feature is the presence of foreground content.

Dataset Structure

dataset/
├── train/
│   ├── motion_0000.mp4 to motion_0449.mp4 (450 videos)
│   ├── text_0000.mp4 to text_0449.mp4 (450 videos)
│   └── metadata.csv
└── test/
    ├── motion_0450.mp4 to motion_0499.mp4 (50 videos)
    ├── text_0450.mp4 to text_0499.mp4 (50 videos)
    └── metadata.csv

Usage

from datasets import load_dataset

# Load dataset
dataset = load_dataset("mukul54/motion-text-classifier")

# Access samples
train_sample = dataset['train'][0]
print(train_sample['label'])  # 'motion_only' or 'text_fg_bg'
print(train_sample['text'])   # word shown (None for motion_only)

Video Parameters

All videos share identical generation parameters:

Parameter	Value
Resolution	960×540
FPS	60
Duration	5.0 seconds
Noise Density	0.5
Speckle Size	1
Speed	2
Direction	Horizontal

Labels

motion_only: Background noise only (no foreground object)
text_fg_bg: Text revealed through opposite motion of foreground/background using the same noise pattern

Key Features

Pure motion detection: Text is invisible in static frames
Controlled experiment: All parameters identical except foreground content
Temporal encoding: Information only available through motion analysis

Use Cases

Testing motion perception in vision models
Temporal feature extraction benchmarks
Zero-shot video understanding
Motion-based object detection

Citation

If you use this dataset, please cite:

@dataset{motion_text_classifier,
  title={Motion-Text Classifier: Speckle Noise Motion Detection Dataset},
  author={Mukul},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/mukul54/motion-text-classifier}
}

License

MIT License