A comprehensive image dataset of American Sign Language hand gestures

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Assistive Technology

Background:

• Existing American Sign Language (ASL) datasets lack sufficient diversity and may not accurately represent alphanumeric characters.
• The distinction between the ASL letter 'O' and the digit '0' is crucial for clear communication but often overlooked in datasets.

Purpose of the Study:

• Introduce ASL-HG, a novel, comprehensive dataset for American Sign Language (ASL) gesture recognition.
• Provide a balanced and diverse dataset to facilitate the development of robust and fair ASL recognition systems.
• Support research in assistive technologies and human-computer interaction for the deaf and speech-impaired community.

Main Methods:

• Collected 36,000 static images across 36 classes (A-Z, 0-9) from 10 volunteers in Bangladesh.
• Ensured balanced representation across subjects, genders, and skin tones, with 100 samples per class.
• Released the dataset in two formats: raw images and MediaPipe-processed hand-segmented crops with a train-test split.

Main Results:

• ASL-HG includes 36,000 images covering the English alphabet and digits.
• The dataset explicitly differentiates the ASL 'O' and '0' signs, enhancing practical communication representation.
• Provided both raw and processed data versions to accommodate diverse research needs.

Conclusions:

• ASL-HG serves as a valuable benchmark for advancing ASL recognition technologies.
• The dataset aims to reduce communication barriers and promote inclusivity for deaf and speech-impaired individuals.
• Facilitates broader research in gesture-based human-computer interaction and the development of effective assistive tools.