Slow FPS on Neural Compute Stick 2

[Joel_E_]

Hi,

I'd like to ask for support specific to the new Movidius MyriadX (Neural Compute Stick 2).

While running the "crossroad_camera_demo.exe" for person detection, the FPS is around 1.68fps on MYRIAD. While on my laptop's CPU it runs around 5fps.

My expectation was, running inference on MYRIAD will increase/improve the frame rate. Could someone confirm if my expectation was wrong?

I am running on Windows 10, and the project is built on VS2017/Debug mode. Please see attached files for more information.

CLI sample:

FPS on Myriad:

FPS on CPU:

If there any information I missed that's why the FPS is slow, I would be happy to hear it.

Thanks in advance!

Joel
 

[Monique_J_Intel]

Hi Joel,

I would highly recommend using person-detection-retail-0013 as it gives roughly about 6 FPS with decent accuracy for the crossroad demo.

Kind Regards,

Monique Jones

[Joel_E_]

Hi Monique,

Thank you. Let me check it and see how it goes.

Regards,

Joel

[Joel_E_]

Hi Monique,

I did use the person-detection-retail-0013 model that you've mentioned, and indeed I got 6-FPS in it. Theoretically if I have 4 pcs. of NCS2, will it give me a 4x boost on performance like 24-FPS? Also will the data being transferred from host to USB be a possible bottleneck when run in parallel?

Regards,

Joel

[Kumar__Shiva]

Hi Joel,

Did you experiment with multiple pieces of NCS2 and notice any performance gain? I did see some videos with raspi connected to multiple NCS2, not sure about the gain though.

Regards,

Shiv

[Hyodo__Katsuya]

I tried NCS2 x 4, but when using Raspberry Pi, the inference speed dramatically improved, but preprocessing and post processing became a bottleneck. As long as it is not high resolution, USB transfer rate rarely becomes a bottleneck.