Intel Movidius Stick 2 vs OAK-D

tarocal

Hey, I am playing around with the movidius Stick 2 and OAK-D.
I tried to run 'face-detection-retail-0005' on both the devices, just to see.

Stick: VM Xubuntu 18.04, USB 3.0
Framerates : 32-35 FPS

OAK-D: VM Xubuntu 18.04, USB 3.0
Framerates : 22-25 FPS

Is such a low framerate for the Stick 2 expected? or am I doing something wrong?
As I understand it, do they not share similar chipset for inferencing?

P.S.: The instructions and examples for the Stick could be better.
P.P.S: I'd like to thank this forum, and its active members and moderators for all their continued help. I try not to reply to the post with a thank you, so if you made it this far, I am talking about you.

GergelySzabolcs

tarocal

Hello,
did you run only inference on both devices, image sent from host like here.
Otherwise if inference is running on RGB camera, the camera handling also contributes to system load, leaving less resources for NN.
The stick does one thing: NN, so it's expected to be somewhat faster.
Regardless, you can try this PR, which reduces the performance gap.
Make sure the NN blob is compiled for optimal resources (6 shaves if cameras are enabled, otherwise 8), although if it's suboptimal a WARNING will be printed.

Brandon

Thanks, Szabi! And thanks for the kind words, @tarocal !

One other note @tarocal , so when used with a fast host, running the images through the host will be faster (as the host is then doing a lot of the work so the Myriad X doesn't have to). But the intention of DepthAI/OAK-D is to allow good/real-time performance with not-fast hosts (and we even have one that works with the ESP32 as a host, here, which is a device with 0.5MB of RAM - that's right, 0.5 MegaBytes of RAM).

So when using something like a Pi as the host, these image manipulations being done on the host, and the latency induced, results in something like a 3x to 8x slow down compared to DepthAI doing them.

This is why Adrian at PyImageSearch sees _8FPS for MobileNetSSD with Pi 3 and NCS2 (see here) whereas with DepthAI we are seeing 50+ FPS for the same network.

Thoughts?

Thanks again,
Brandon