How to improve the performance of the object detection example in the Android sample?

[libofei2004]

I run the object detection example https://ai.google.dev/edge/mediapipe/solutions/vision/object_detector?hl=zh-cn%2Cincludes in MI10 pro, my model is EfficientDet-Lite0 (int8) downloaded from https://ai.google.dev/edge/mediapipe/solutions/vision/object_detector?hl=zh-cn%2Cincludes.

As the page says, the model's performance can reach within 29.31 milliseconds in CPU mode, but in reality, the fastest I can run is about 60ms, usually only around 80 or 90 ms, and there is not much difference between using GPU mode and CPU mode.

How to improve the performance of the object detection example in the Android sample?