We implement a Filipino Sign Language Recognition model for both static and dynamic hand gestures using MediaPipe extracted landmarks. The files in this repository are separated into three categories: video processing, model processing, and web application.

First install pytorch 1.13+ and other 3rd party dependencies.

conda create --name fsl-translator python=3.9 -y
conda activate fsl-translator

pip install -r requirements.txt

Under video processing:

1. arrange_files.ipynb is a notebook that manipulates directories and files for convenience of the user. Example is by concatenating old video files to newer ones, and concatenating MP_Data folders together etc.
2. video_capture.ipynb captures 30 frame videos N times for convenience of the user. This notebook explains the directory format followed by the project.
3. video_preprocess.ipynb: In order to implement a Senyas, we first preprocess the videos via squeezing/augmenting the video inputs into 30 frames.
4. video_landmark_extraction.ipynb: The video inputs are then sent into OpenCV to capture frames and into the MediaPipe Holistic pipeline for landmark extraction for each frame.

Under model processing:

1. model_training.ipynb: The extracted landmarks are saved in MP_Data folder which is used as input to the model architecture. This model can be evaluated in this notebook using a confusion matrix and accuracy score.
2. model_testing.ipynb: This notebook allows the user to test the saved model in model_training.ipynb.

You can visit the official website at Official Website.

You can also run the web application locally. Check the official documentation for more details: Official Documentation

Senyas: FSL Translator code and model weights are released under the CC-BY-NC 4.0 license. See LICENSE for additional details.

If you find this repository useful, please consider giving a star ⭐ and citation