AI4ALL_ASL_AI

---

MediaPipe.ipynb

Phase 1:

• initially used resnet-50 to transfer learn ASL alphabet using 1200 collected image per alphabet
• Peak accuracy: 77% (low accuracy is likely due to complexity/lack of localization of hand as it takes input of whole image.

Phase 2:

• Instead used Mediapipe to extract 63 points (21 points of x, y, z of a hand joints)
• Then used Multi-layer perceptron to classify
• Peak accuracy 98.2%, performed very good even in low quality camera as it only need 63 point from mediapipe to be properly given.