Difference in performance between sample face mesh app and web demo.

[philiprkk]

It seems like there's a significant difference in responsiveness between the face landmarker app for Android and the web demo. Is there a reason for this? I'd like to develop an android app with the best real-time responsiveness as possible, but it seems like there's some sort of buffer playing catch-up sometimes, even when using asynchronous detection, which should solve this issue. This is on an Xperia 5 II with both using GPU delegate.

https://github.com/google-ai-edge/mediapipe-samples/assets/57251565/a637cb08-952a-4e55-8d7d-799df107bc94