mohibkhansherwani/DynamicWordLevelPakistanSignLanguageDataset

---
license: apache-2.0
task_categories:
- tabular-classification
- feature-extraction
language:
- ur
- en
tags:
- sign-language
- mediapipe
- landmarks
- pakistan
- gesture-recognition
pretty_name: Dynamic Word Level Pakistan Sign Language (PSL) Dataset
size_categories:
- 1K<n<10K
---

# Dynamic Word Level Pakistan Sign Language (PSL) Dataset
## Overview
This dataset contains **MediaPipe hand landmark sequences** for 60+ words in Pakistan Sign Language (PSL). It is designed to support research into dynamic, word-level gesture recognition. Unlike existing datasets that focus on static finger spelling or small vocabularies, this project provides a high-quality, research-ready landmark collection for complex sign language translation.

**Kaggle Link** https://www.kaggle.com/datasets/mohib123456/dynamic-word-level-pakistan-sign-language-dataset/data

**Github Project Link** https://github.com/MohibUllahKhanSherwani/SignSpeak_FYP
### Dataset Statistics
- **Total Signs:** 60+ unique word classes.
- **Samples per Sign:** Each sign is performed 70 times (Combined).
- **Sequence Length:** Each gesture is fixed at 60 frames for temporal consistency.
- **Format:** Landmark coordinate data (X, Y, Z) extracted via MediaPipe.
## Dataset Structure & Generalization
The data is organized into two primary subsets to ensure model robustness across different hardware and camera types:
1. **MP_Data:** contains 50 samples per sign recorded using standard webcams and fixed desktop cameras.
2. **MP_Data_mobile:** contains 20 samples per sign recorded using mobile phone cameras to introduce varied angles, motion blur, and lighting.
By training on both sets, models are less likely to overfit to a specific lens or environment, making them more suitable for real-world mobile applications like **SignSpeak**.
## Motivation & Rationale
Most existing PSL datasets are limited to static signs or small vocabularies. We collected this word-level dataset specifically for the SignSpeak project.
- Privacy: No raw video recordings were saved to protect participant identity. Only anonymized MediaPipe landmarks were kept.
- Scale: Large vocabulary (60+) to move beyond simple finger-spelling
## Usage
This data is ideal for training sequence-based architectures like **LSTMs, GRUs, or Transformers** for temporal classification. The primary use case is for building real-time sign language translation tools for the Deaf community in Pakistan.
### Loading the Data
For details on how to use the dataset and sample loading and training script please visit: https://www.kaggle.com/datasets/mohib123456/dynamic-word-level-pakistan-sign-language-dataset/data
## Authors
This dataset was curated by the **SignSpeak** team as part of our Final Year Project at **COMSATS University Islamabad, Abbottabad Campus**.
- **Main Author:** Mohib Ullah Khan Sherwani
- **Repository:** https://github.com/MohibUllahKhanSherwani/SignSpeak-FYP
## License
This dataset is released under the **Apache License 2.0**. You are free to use, modify, and distribute this data for both research and commercial purposes, provided that attribution is given to the authors.