Teach the camera your own classes

Define two or three classes — your mug, your hand, nothing — capture fifteen examples of each with the camera and train on top of a frozen large model. Two seconds later you have your own classifier. This, transfer learning, is how most real vision problems get solved.