Best Practices

The steps described in previous sections are usually enough in order get you to the level of a working demo. You may, however, not be satisfied with the success rate: gestures are recognized most of the time but that might not be enough for a real world solution.

The main reason for this is the inconsistent nature of the way gestures are normally performed. The characteristics of gestures are influenced by many factors such as: how fast we do it, the size of our arm, your posture, etc. Consequently, even a simple movement, such as raising our hand may vary significantly.

Depending on your use case, achieving good results require that you record multiple different variants of the same gesture. For example, if you want to know when people are raising their hand, you will have to be able to recognize different variants of that gesture, such as: raising hand quickly, raising hand slowly, man with long arm, woman with short are, while standing, while sitting, etc.

Each of these variants is considered a different gesture by the algorithms. Depending on your actual use case, your app should know how to test a gesture with multiple algorithms and generalize the gesture type. For example, using an or condition may be sufficient for generalizing gestures.