Overview
Engineered a YOLOv5 pipeline to classify 12 Naruto hand signs from live webcam frames and drive real-time jutsu logic. Rebuilt the earlier Mediapipe + RandomForest approach to improve robustness, and converted datasets to annotated YAML with Roboflow for scalable training.
Highlights
- Augmented training data with rotation, hue jittering, and frame-interval captures in OpenCV.
- Reduced overfitting through randomized variance injection and broader spatial detection.