Gesture Recognition for Video Game Controllers
Jonatan Ebenholm
Compares two approaches to real-time gesture recognition for game input: a custom CNN trained on the HaGRID dataset plus self-collected images, versus a Feed Forward Neural Network operating on 21 3D hand landmarks extracted with MediaPipe. The MediaPipe + FFNN hybrid reached 100% validation accuracy with sub-second per-epoch training, outperforming the CNN on both accuracy and rotation robustness.