Sign4all: a Spanish Sign Language dataset.

Area of Science:

• Computer Science
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Sign Language Recognition (SLR) is crucial for inclusive technology.
• Existing datasets face challenges like data sparsity and right-handed bias.
• There is a need for comprehensive datasets for specific sign languages like Spanish Sign Language (LSE).

Purpose of the Study:

• Introduce Sign4all, a novel dataset for Isolated Sign Language Recognition (ISLR) in Spanish Sign Language (LSE).
• Address data sparsity and handedness bias in current SLR datasets.
• Facilitate the development of robust deep learning models for LSE.

Main Methods:

• Collected 7,756 high-resolution RGB videos and skeletal keypoints for 24 LSE signs (catering vocabulary).
• Implemented a high-density approach with an average of 323 samples per sign.
• Ensured handedness balance (equal left/right-handed signs) and applied manual segmentation, temporal, and spatial normalization.
• Formatted data in AVI (video) and HDF5 (keypoints) for compatibility with deep learning frameworks.

Main Results:

• The Sign4all dataset offers an average of 323 samples per sign, significantly reducing data sparsity.
• Achieved handedness balance, crucial for developing models invariant to sign handedness.
• Technical validation using Transformer and skeletal models confirmed dataset integrity.
• Demonstrated the necessity of pre-computed augmentation splits for effective model training.

Conclusions:

• Sign4all provides a valuable, high-density, and balanced resource for Spanish Sign Language Recognition research.
• The dataset's design directly addresses limitations of previous SLR datasets.
• Sign4all supports the advancement of inclusive human-machine interaction technologies for the deaf and hard-of-hearing community.