Skip to content

samueldavidbryan/MachineLearning-FingerspellingDetection

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 

Repository files navigation

Translating ASL Fingerspelling

Project Site

Project site with full overview of our efforts.

Background

This repository focuses on translating American Sign Language (ASL) fingerspelling into text using deep learning, particularly Long-Short-Term-Memory (LSTM) networks. Voice recognition is prevalent, but gesture-to-text technology is still underdeveloped. ASL native signers are much faster at fingerspelling than typing on a mobile device. The project is based on a Kaggle competition with a 180GB dataset of ASL phrases, normalized using Mediapipes.

Solution

The project utilizes LSTM networks for processing sequential frames representing ASL phrases. The dataset contains 123 parquet files with labeled ground truth. The files include body, hand, and face coordinates, mapped with coordinates extracted from video frames.

How to Run

We recommend you run the notebook with the Kaggle environment since the data set is very large, around 180GB and is already hosted on Kaggle. Open a new Kaggle cloud-based notebook and import this notebook to that environment. Select GPU as the hardware accelerator option. To connect to the data source, In the "Data" section on the right hand side, click on "Add Data" and search the keyword "Google - American Sign Language Fingerspelling Recognition" and click on "add competition".

Alternatively, you can "Copy & Edit" the our notebook at https://www.kaggle.com/code/minhhangr/gompei-ninjas-asl-fingerspelling-with-lstm to get your own copy of our notebook.

Resources

Team members

  1. Minh-Hang Radetsky
  2. Sam Bryan
  3. Adish Jain
  4. Sireesha Bachu

About

Summer 2023 WPI CS539 Machine Learning Final Project (Source Code): Detecting American Sign Language (ASL) Fingerspelled Phrases

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published