How to achieve realtime inference?

Question

John1231983 opened this issue 5 years ago · 7 comments

Hi, the inference code runs very fast. Could you tell me any trick to obtain the realtime inference? Does it run on GPU or CPU?

yes

Answer 1 · 2020-01-02T02:45:16.000Z

Android app running on arm cpu with ncnn. if you want design a realtime landmark detector, a lightweight network is needed.

Answer 2 · 2020-01-02T02:50:36.000Z

lightweight network

I am using mobilenet v2 with alpha 0.25 as the paper mention. Do you use same thing or customize the mobilenet v2? Any suggestion?

Answer 3 · 2020-01-02T02:58:00.000Z

mobilenet v2 is ok

Answer 4 · 2020-01-02T03:47:27.000Z

Thanks. But I found that the speed is still slow. I used canvas to draw the landmark points. Is it fine?

Answer 5 · 2020-01-02T03:49:40.000Z

If you running on android or ios devices, draw points based on opengl.

Answer 6 · 2020-01-02T03:54:56.000Z

opengl is faster than canvas API function. Is it right? Thanks so much