Mexican Sign Language Glosses (Dynamic Signs)

Mexican Sign Language Glosses (Dynamic Signs)

Overview

The Mexican Sign Language (LSM) Glosses Dataset is a specialized collection of 121 dynamic signs designed to facilitate research in sign language recognition within healthcare contexts. This dataset covers a comprehensive lexicon tailored for medical emergencies, accident scenarios, common symptoms, and polite expressions essential for effective patient-provider interactions.

Data Representation

Format: High-definition video files in .mp4 format.

Content: Each video represents a single gloss (dynamic sign) executed from start to finish.

Environment: All recordings were captured in a controlled environment featuring optimized lighting conditions and a uniform green background to simplify background subtraction and keypoint extraction.

Perspective: All signs were recorded from a frontal view to provide a direct representation of hand shapes and body movements.

Data Collection and Participants

The dataset includes contributions from 12 participants, offering a balanced mix of expertise:

Experts: Two key subjects were included to ensure linguistic accuracy—a professional LSM interpreter (Subject 0) and a Deaf signer (Subject 11).

Non-Experts: Ten subjects (non-signers) were included to provide variability in sign execution, which is crucial for training robust machine learning models.

Repetitions: Each subject performed each of the 121 signs once, resulting in a diverse set of executions for each gloss.

Repository Structure

The data is organized by gloss name to simplify the training pipeline. Each folder contains the video files corresponding to that specific sign:

Mexican Sign Language Glosses/

├── <Gloss_Name>/ │    ├── <Gloss_Name>_<SubjectID>.mp4 │    └── ... └── ...

Naming Convention: <Gloss>_<SubjectID>.mp4

Gloss: The specific medical or polite expression (e.g., "Dolor", "Ayuda", "Hospital").

SubjectID: Numerical identifier (0–11).

Data Partitioning

As this is a primary corpus, no predefined train/test split is provided. This allows researchers to implement their own validation strategies, such as Leave-One-Subject-Out cross-validation, which is highly recommended given the participant-dependent nature of sign language.