Dataset Documentation

Overview

This dataset was created to support the development of a real-time gesture recognition system for Sri Lankan Sign Language (SLSL).

The dataset focuses on a subset of 10 static hand gesture letters: A, E, I, O, U, L, N, R, S, T.

As no suitable public dataset was available for this specific use case, a custom dataset was collected and curated.

Dataset Structure

The dataset is organised into class-based directories, where each folder represents a gesture class (letter).

dataset/
├── A/
│   ├── A001.jpg
│   ├── A002.jpg
├── E/
│   ├── E001.jpg
│   ├── E002.jpg
├── I/
│   ├── I001.jpg
│   ├── I002.jpg
├── O/
│   ├── O001.jpg
│   ├── O002.jpg

Each folder contains images that correspond to a single gesture.

Dataset Composition

  • Number of letters: 10

  • Letters included: A, E, I, O, U, L, N, R, S, T

  • Total number of images: 456

The dataset is approximately balanced with each class containing a similar number of images.

Dataset Summary

Letter

Image Count

A

45

E

45

I

46

O

45

U

44

L

47

N

47

R

47

S

45

T

45

Data Collection

Images were collected using standard smartphone cameras.

To improve dataset diversity and model generalisation, images were captured under varying conditions:

  • Lighting: natural light, indoor lighting, and low-light environments

  • Backgrounds: plain and cluttered backgrounds

  • Angles: slight variations in hand orientation

  • Participants: multiple individuals with different hand sizes and skin tones

File Naming

All images follow a consistent naming format:

LETTER+NUMBER.jpg

  • LETTER corresponds to the class label

  • NUMBER is a sequential identifier

Examples:

A001.jpg
E012.jpg

Images were converted to .jpg format using ImageMagick.

Usage

This dataset is intended for use in image classification tasks related to sign language gesture recognition.

It can be used to:

  • Train machine learning models for gesture classification

  • Evaluate model performance on static hand gestures

  • Develop and test real-time sign language recognition systems

Ethical Considerations

  • No personally identifiable information is included

  • Only hand gesture images were collected

  • Participation was voluntary

  • Data is used strictly for academic purposes

Limitations

  • Dataset size is relatively small (456 images total)

  • Only a subset of letters is included

  • Static gestures only

  • Slight class imbalance between letters

  • Environmental conditions may affect model performance

  • Regional variations in SLSL are not fully represented

Future Improvements

  • Expand dataset to include the full alphabet

  • Increase dataset size and diversity

  • Include dynamic hand gestures

  • Collect data from multiple regions

  • Improve variation in environmental conditions