Today I worked on an atom with NCS offloading project. I am baffled with
a situation when running interactive_face_detection_demo sample in OpenVino R5.
When CPU was doing inferencing, it took only 34.05ms for CPU to cap/render,
But when NCS 2 was offloading CPU, the CPU took 379ms to cap/render each frame!
Does anyone know why? And how would I fix the bottleneck?
NCS 2 inferencing
Thank you in advance,
Dear Henry, Lin,
Can you kindly retry your experiments with the most recent version of OpenVino - 2019.R.1.0.1 ?
According to Interactive Face Detection Documentation there are a few knobs you can play with to get better performance (such as enabling async mode, dynamic batching). But you can find out precisely what is happening by running the sample with the -pc argument which will give you a per-layer performance report.
Hope it helps,