Eye-Rubbing Detection Using a Smartwatch

This repository contains the dataset used in the paper:

Eye-Rubbing Detection Using a Smartwatch: A feasibility study demonstrated high accuracy with machine learning methods based on Transformer

Paper in TVST, Sep 2024

Code in GitHub

Published in Translational Vision Science & Technology (TVST)

Dataset Description

This dataset was collected for the purpose of developing a machine learning model to detect eye-rubbing and other hand-face interactions using smartwatch sensors. The data was collected using an Apple Watch and includes various sensor readings such as accelerometer, gyroscope, and orientation data. For more details about the dataset and the research behind this project, please refer to the paper linked above.

Dataset Statistics

Total Users: 50
Total Signals: 11,531
Total Duration: 18 hours 20 minutes
Classes: 8 main classes + 1 "Nothing" class

Data Collection Methods

Automatic Labeling Setup:
- Used OpenPifPaf computer vision software for motion detection
- Participants completed 20-minute sessions (5 sets of 4 minutes each)
- Resulted in signals of variable length
Manual Labeling Setup:
- Participants performed specific hand-face interactions
- Each action was recorded for 3 seconds
- Resulted in fixed-length sequences

Class Distribution

Eye Rubbing: 1,618 sequences
Eye Touching: 660 sequences
Glasses Readjusting: 638 sequences
Eating: 635 sequences
Make Up (Application + Removal): 687 sequences
Hair Combing: 705 sequences
Skin Scratching: 755 sequences
Teeth Brushing: 970 sequences
Nothing: 4,863 sequences

Data Format

19 features are recorded for each time step:
- Raw Accelerometer Data (x, y, z)
- Processed Device-Motion Data:
  - Yaw, Roll, Pitch
  - Rotation Rate (x, y, z)
  - User Acceleration (x, y, z)
  - Quaternion (x, y, z, w)
  - Gravity (x, y, z)

Usage

This dataset is suitable for developing and evaluating machine learning models for hand-face interaction detection, particularly eye-rubbing detection. It can be used for both supervised learning tasks and unsupervised pre-training.

Ethical Considerations

The data was collected with participant consent. However, researchers should be mindful of potential biases in the dataset and the ethical implications of developing technology for monitoring personal behaviors.

Citation

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

[TO BE ADDED]