ASL Fingerspelling Classifier

Overview

Which approach recognizes American Sign Language fingerspelling best for the compute it costs? This project compares three ways of classifying the 26 ASL letters on the Kaggle ASL Alphabet dataset.

What I built

• CNN: fine-tuned a VGG network on raw hand images.
• Feature matching: classical keypoint matching with ORB descriptors.
• Skeletal: extracted 21 MediaPipe hand landmarks (63 features) per frame and trained a lightweight classifier on them.
• Deployed the skeletal model as a real-time webcam translator.