How to improve recognition speed

[zhou123aw]

Why is it that when I perform hand landmarks detection , the sensitivity of recognition is not high and the speed is slow?
Is it because of. Net? I am using the Python version with fast speed.

[sr229]

You will need to lower the resolution input for Mediapipe, if you put it anything above 640x480, tracking slows down to a crawling halt.

[zhou123aw]

You will need to lower the resolution input for Mediapipe, if you put it anything above 640x480, tracking slows down to a crawling halt.

@sr229 I have tried to reduce the resolution to 320 * 240, but the problem still exists. It seems that the recognition effect on the back of the hand is relatively poor. When I extend the palm part, the effect is relatively good. Didn't I add a model of the back of the hand?

[zhou123aw]

@sr229 The recognition effect seems to be related to the proportion of my palms exposed. For example, when I gesture with the number 1, the recognition effect is poor due to the less palms exposed. However, when I gesture with the number 5 or 4, the recognition speed is very fast.

[sr229]

Have you tried performing a benchmark against native mediapipe to see if the speed is comparable? There's nothing I can do mostly about if it's an upstream problem.