| --- |
| 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. |
