Intel® Distribution of OpenVINO™ Toolkit
Community assistance about the Intel® Distribution of OpenVINO™ toolkit, OpenCV, and all aspects of computer vision-related on Intel® platforms.
6506 Discussions

video_objects_scalable slows down when increasing ncs number

idata
Employee
1,057 Views

Hi,

 

I'm using ncs v1 with a pi 3+. With one ncs I have 8fps, 2: 15 fps. But with 3, the fps number doesn't change, and with 4 ncs, the fps slows down to 7 .
0 Kudos
7 Replies
idata
Employee
798 Views

@spi100

 

Are you using Python?

 

If you are using Python, the performance of MultiThread will be exacerbated by GIL (Global Interpreter Lock).

 

Performance improves when using MultiProcess.

 

Sample of MultiProcess + MultiThread + MultiStick + MultiGraph + etc…

 

https://github.com/PINTO0309/MobileNet-SSD-RealSense.git
0 Kudos
idata
Employee
798 Views

@PINTO yes, I do. Thanks for the tip, I'm going to have a look.

 

Regards
0 Kudos
idata
Employee
798 Views

@PINTO I've read your git hub page, a clear and pretty good job. Now I've a better understanding of what you explained in your previous post.

 

If I want to reach the 25 fps real time, at first, to avoid the GIL issue, I can write my code in C or C++.

 

Perphaps also using an aaeon card and a realsense camera is likely a better solution than using a simple pi 3+ with a piCam.
0 Kudos
idata
Employee
798 Views

@spi100

 

 

If I want to reach the 25 fps real time

 

 

I realized the performance of over 30 FPS over NCS2 x2 + RaspberryPi3.

 

 

Perphaps also using an aaeon card and a realsense camera is likely a better solution than using a simple pi 3+ with a piCam.

 

 

There are points to be noted.

 

The minimum resolution of RealSense is 640x480.

 

Inference by NCS or NCS 2 is very fast, but it is a very bottleneck because of the high load (640x480) on the preprocessing and post-processing CPUs.

 

My benchmark understands that pre-processing and post-processing abilities are overwhelmingly deficient in ARM CPUs.

 

Perhaps aaeon is a very good choice.

 

If you plan to introduce OpenVINO, I encourage you to refer to this benchmark.

 

When using Intel's CPU, OpenVINO has been optimized considerably.

 

https://ncsforum.movidius.com/discussion/1329/lattepanda-alpha-openvino-cpu-core-m3-vs-ncs1-vs-ncs2-performance-comparison
0 Kudos
idata
Employee
798 Views

@PINTO Hi, I've tried to run your sample

 

python3 SingleStickSSDwithUSBCamera_OpenVINO_NCS2.py -mod 1 -wd 640 -ht

 

480

 

after installing openVino on my pi3+.

 

I haven't any error message displayed on the console but no window is displayed on the screen.

 

Have you got any idea ?

 

regards

0 Kudos
idata
Employee
798 Views

The call to cv2.VideoCapture(0) is not ended. I use 'modprobe bcm2835_v4l2 ' to simulate an usb camera with a piCam. It could be the root cause of my issue ?

0 Kudos
idata
Employee
798 Views
0 Kudos
Reply