ISLR101: Iranian Word-Level Sign Language Recognition Dataset

Hossein Ranjbar¹, Alireza Taheri²

¹ Department of Computational Linguistics, University of Zurich, Zurich, Switzerland
² Department of Mechanical Engineering, Sharif University of Technology, Tehran, Iran

---

Introduction

we introduce ISLR101 dataset, the first publicly available Iranian Sign Language dataset for isolated sign language recognition. This comprehensive dataset includes 4,614 videos covering 101 distinct signs, recorded from 10 different signers, along with pose information extracted using OpenPose. We establish visual appearance-based and skeleton-based frameworks as baseline models, thoroughly training and evaluating them on ISLR101 to demonstrate their effectiveness.

This repository provides a PyTorch-based implementation of Skeleton-based sign language recognition.

Instalation

1. Clone this repository

git clone https://github.com/HoseinRanjbar/ISLR101.git
cd ISLR101

2. Download ISLR101 pose data

mkdir data
cd data
wget https://drive.google.com/uc?export=download&id=1mqWgZJ7mJZEDyuK5lixC4g4ZUKa1ZKme
wget https://drive.google.com/uc?export=download&id=1Q1Y1noTdG0pJSLecLqZnvt_fnNbs306I
cd ..

3. Install dependent packages

pip install -r requirements.txt

Usage

1. Test

To test the model on the ISLR101 dataset, use the following command:

• ttr configuration:

./scripts/test_ttr.sh

• str configuration:

./scripts/test_str.sh

• sttr1s configuration:

./scripts/test_sttr1s.sh

2. Training

To train the model on the ISLR101 dataset, use the following command:

• ttr configuration:

./scripts/train_ttr.sh

• str configuration:

./scripts/train_str.sh

• sttr1s configuration:

./scripts/train_sttr1s.sh