For the architecture of our gesture interface, we emphasized reusability and integration between components. We divided the architecture into five layers, which are (1) Hardware interface (2) Segmentation/tracking (3) Classification (4) Modeling and (5) Gesture recognition. There are six steps to the hand regions classification method, which are (1) Synthetic hand poses creation (2) Decision trees training using GPU (3) Hand segmentation (4) Decision forest classifier using GPU (5) Post-processing and (6) Estimating joint position.

Step 1 and 2 are performed offline to generate decision trees and step 3 to 6 are executed online to classify the input image in real-time.