Hi, I'm a starter and trying to learn how to customize or modify my own mediapipe line. I used Neural Networks to train landmarks which extract from mediapipe. Is there any way I can put my trained model back to mediapipe to implementing real-time gesture recognition？Thanks for your help.

#546 adds a lengthy pattern that gets the inverse of a state before composing it. Instead we should have an inverse composition (i.e. subtraction) function implemented between states for brevity.

@gilwoolee

Hi, first of all thanks for sharing your work, I really appreciate it!

Would it be possible to add some documentation on how to compile the project and what the dependencies are?