We tried a lot of different strategies in—

pre-made models

hand and face detection separately

training the algo yourself

and it felt like the last one was the strongest (which is what we released). The one we built is as strong as the training you give it—it shouldn't be going off if your hand is at all in the frame.