manojkumarcs/INCLUDE_Dataset

🤟 INCLUDE: A Large-Scale Dataset for Indian Sign Language Recognition

A comprehensive video dataset for Indian Sign Language Recognition

🌐 Computer Vision  •  🧠 Deep Learning  •  🎥 Video Recognition  •  🤟 Sign Language Recognition

---

👋 Welcome

Hello and welcome!

Thank you for your interest in INCLUDE: A Large-Scale Dataset for Indian Sign Language Recognition.

We are pleased to make this dataset available to the research community to support the development, evaluation, and benchmarking of Indian Sign Language Recognition (SLR) systems.

🤟 Making Indian Sign Language research more accessible — one sign, one video, and one model at a time.

INCLUDE provides a diverse collection of short ISL signing videos recorded with experienced signers. The dataset is intended to support research in computer vision, deep learning, video-based action recognition, and sign language understanding.

We hope INCLUDE will contribute to:

• 🔬 Reproducible SLR research
• 📊 Standardized benchmarking
• 🧠 Development of robust deep learning models
• 🎥 Video-based sign recognition
• ♿ Accessibility-focused technologies
• 🇮🇳 Advancement of Indian Sign Language research

---

📊 Dataset at a Glance

🎥 Video Sequences	🤟 Word Signs	🗂️ Categories	⚡ INCLUDE-50
4,292	263	15	50

🔎 Dataset Structure

                         ┌─────────────────────────┐
                         │        INCLUDE          │
                         │  Indian Sign Language   │
                         │        Dataset          │
                         └────────────┬────────────┘
                                      │
                 ┌────────────────────┴────────────────────┐
                 │                                         │
        ┌────────▼─────────┐                     ┌─────────▼────────┐
        │  INCLUDE Dataset │                     │    INCLUDE-50    │
        │                   │                     │                  │
        │  4,292 Videos     │                     │   50 Word Signs  │
        │  263 Word Signs   │                     │   Rapid Evaluation│
        │  15 Categories    │                     │   & Tuning       │
        └───────────────────┘                     └──────────────────┘

---

📝 Description

INCLUDE (Indian Sign Language Dataset) is a comprehensive video-based dataset developed for Indian Sign Language (ISL) Recognition research.

The dataset contains:

🎥 4,292 short video sequences 🤟 263 distinct word signs 🗂️ 15 diverse word categories

The videos feature experienced Indian Sign Language practitioners, enabling the dataset to capture variations in signing performance while providing conditions that closely resemble natural and realistic signing environments.

This makes INCLUDE suitable for developing, evaluating, and benchmarking Sign Language Recognition (SLR) models under practical conditions.

🚀 INCLUDE-50

To facilitate rapid experimentation and evaluation, we additionally provide INCLUDE-50, a representative subset containing 50 selected word signs distributed across the different word categories.

INCLUDE-50 can be particularly useful for:

• ⚡ Rapid model prototyping
• 🎯 Hyperparameter tuning
• 🧪 Initial experimentation
• 📈 Comparative evaluation
• 🔬 Ablation studies
• 🧠 Model development

Researchers can therefore begin experimentation with INCLUDE-50 before moving to the complete INCLUDE dataset.

---

🗂️ Dataset Organization

For ease of download, storage, and hosting, the complete dataset has been divided into multiple compressed archive files at the word-category level.

Larger categories are further divided into multiple parts to make downloading and managing the dataset more convenient.

📦 Dataset Distribution

The following chart shows the distribution of the 43 compressed archive parts across the 15 word categories:

pie title INCLUDE Dataset Archive Distribution
    "Adjectives" : 8
    "Animals" : 2
    "Clothes" : 2
    "Colours" : 2
    "Days and Time" : 3
    "Electronics" : 2
    "Greetings" : 2
    "Home" : 4
    "Jobs" : 1
    "Transportation" : 2
    "People" : 5
    "Places" : 4
    "Pronouns" : 2
    "Seasons" : 1
    "Society" : 3

📋 Categories

#	Category	📦 Parts
01	📝 Adjectives	8
02	🐾 Animals	2
03	👕 Clothes	2
04	🎨 Colours	2
05	🗓️ Days and Time	3
06	💻 Electronics	2
07	👋 Greetings	2
08	🏠 Home	4
09	👨‍💼 Jobs	1
10	🚗 Means of Transportation	2
11	👥 People	5
12	📍 Places	4
13	👤 Pronouns	2
14	🌦️ Seasons	1
15	🏙️ Society	3
	Total	43 Parts

📦 43 compressed archive parts are distributed across 15 word categories.

This category-level organization also enables researchers to download and experiment with specific categories without necessarily downloading the entire dataset.

---

📥 Dataset Preparation

1️⃣ Download

To access the complete INCLUDE dataset, download all .zip files provided in this repository.

Download
   │
   ├── Adjectives
   ├── Animals
   ├── Clothes
   ├── Colours
   ├── Days and Time
   ├── Electronics
   ├── Greetings
   ├── Home
   ├── Jobs
   ├── Transportation
   ├── People
   ├── Places
   ├── Pronouns
   ├── Seasons
   └── Society

2️⃣ Extract

Select all downloaded .zip files and extract them into a single specified directory.

After extraction, the videos will be organized according to their respective categories.

INCLUDE/
│
├── Adjectives/
├── Animals/
├── Clothes/
├── Colours/
├── Days and Time/
├── Electronics/
├── Greetings/
├── Home/
├── Jobs/
├── Means of Transportation/
├── People/
├── Places/
├── Pronouns/
├── Seasons/
├── Society/
│
└── Train_Test_Split/

3️⃣ Train/Test Splits

A dedicated Train_Test_Split folder is included with the dataset.

It contains CSV files defining the training and testing partitions for:

• 🤟 INCLUDE
• ⚡ INCLUDE-50

The provided splits can be used directly to reproduce the recommended experimental setup.

⚠️ Recommendation: Researchers are encouraged to use the provided train/test split files when comparing different SLR approaches to ensure consistent and reproducible evaluation.

---

🔄 Experimental Workflow

flowchart LR
    A[📥 Download Dataset] --> B[📦 Extract ZIP Files]
    B --> C[🗂️ Category-wise Videos]
    C --> D{Dataset Selection}
    D -->|Full Dataset| E[🤟 INCLUDE<br/>263 Signs]
    D -->|Rapid Evaluation| F[⚡ INCLUDE-50<br/>50 Signs]
    E --> G[📄 Train/Test CSV]
    F --> G
    G --> H[🧠 Train SLR Model]
    H --> I[📊 Evaluate Performance]

---

🧪 Recommended Usage

⚡ Quick Experimentation

Start with INCLUDE-50 when you want to:

Model Development
       ↓
Rapid Training
       ↓
Hyperparameter Tuning
       ↓
Ablation Studies
       ↓
Initial Benchmarking

🚀 Full Evaluation

For comprehensive experiments, use the complete:

INCLUDE → 4,292 videos → 263 signs → 15 categories

This provides a broader benchmark for evaluating the generalization and robustness of SLR models.

---

🔬 Research Applications

INCLUDE can support research in several areas:

Area	Potential Application
🧠 Deep Learning	CNN, RNN, LSTM, Transformer-based SLR
🎥 Video Recognition	Temporal action and gesture recognition
👁️ Computer Vision	Visual representation learning
🤟 SLR	Indian Sign Language recognition
⚡ Efficient AI	Lightweight and real-time SLR
🔬 Benchmarking	Comparative evaluation of recognition models
🧪 Experimentation	Hyperparameter and ablation studies
♿ Accessibility	Assistive communication technologies

---

📈 Why INCLUDE?

                ┌──────────────────────────────┐
                │          INCLUDE             │
                └──────────────┬───────────────┘
                               │
          ┌────────────────────┼────────────────────┐
          ▼                    ▼                    ▼
     🎥 Diverse            🤟 Multiple          🗂️ Category
       Videos                Signs               Diversity
          │                    │                    │
          └────────────────────┼────────────────────┘
                               ▼
                    🧠 Robust SLR Research
                               │
                               ▼
                    📊 Reproducible Evaluation
                               │
                               ▼
                    ♿ Accessible Technology

---

📌 Dataset Statistics

🎥 4,292

Short Video Sequences

🤟 263

Word Signs

🗂️ 15

Word Categories

⚡ 50

INCLUDE-50 Signs

📦 43

Archive Parts

---

📚 Reproducibility

To promote reproducible research, INCLUDE provides predefined training and testing splits.

Researchers are encouraged to:

1. Download the required dataset archives.
2. Extract the archives into a common directory.
3. Use the provided Train_Test_Split CSV files.
4. Train the selected SLR architecture.
5. Evaluate the model using the corresponding test split.
6. Report results using clearly defined evaluation metrics.

This setup enables more meaningful comparison between different approaches.

---

🙏 Thank You

We sincerely hope that the INCLUDE and INCLUDE-50 datasets will serve as valuable resources for research in Indian Sign Language Recognition.

We hope these datasets contribute to the development, evaluation, and benchmarking of robust, efficient, and practical SLR systems, while encouraging further research in Indian Sign Language, computer vision, deep learning, and accessibility technologies.

Your research and applications can help move the field forward.

🤟 Every sign matters. Every dataset contributes. Every model brings us one step closer to accessible communication.

We appreciate your interest in INCLUDE and look forward to seeing the research, applications, and innovations enabled by these datasets.

🌟 Happy Researching! 🌟

Thank you for being a part of the INCLUDE research community!!!

🤟 🇮🇳 🧠 🎥 🔬