Image Based American Sign Language (ASL) Recognition with PyTorch

This code was originally written for a student project at the Technical University Munich. The goal was to train a state of the art (2017) deep CNN to classify RGB images showing finger spelling signs. We achieved an accuracy of 83.6% on a subset of speaker E from the ASL Finger Spelling Dataset with a DenseNet 50. The model was trained on speaker A to D as well as on images from RWTH Finger Spelling Database.