ConfuseIous / ASLearn

My winning submission for the WWDC 2022 Swift Student Challenge.
8 stars 0 forks source link
asl coreml swift swiftui uikit vision

ASLearn

ASLearn has been accepted by Apple for the 2022 Swift Student Challenge!!!

ASLearn aims to teach American Sign Language in an interactive manner with the help of Machine Learning.

Please note that this Playground has been designed for, tested on, and must be run on iPad.

Info

This project has multiple phases.

Training Phase

  1. I obtained the ASL Alphabet dataset from Kaggle, sourced under the GPL 2 License.
    1. I found and modified a Python script (used under the MIT License) leveraging the deeplabv3_resnet101 Machine Learning Model (used under the Apache License) to remove the background of the provided image and modified it to remove the background of every image in the dataset.
    2. I trained an Image Classifier using CreateML and the dataset sans background.
      • I also built my own Image Classifier using Tensorflow but achieved higher accuracy with CreateML and thus used that instead. I do plan to refine my Image Classifier in the future and make ASLearn available on the App Store.

        Actual Usage

    3. ASLearn takes a picture of the user's hand gesture
    4. The hand gesture is provided to DeepLabV3 and a result without a background is obtained.
    5. The result is provided to ASL_Classifier (described in Step 3 of the Training Phase) and a prediction is received.

Credits