Yolov3 inference on VPU cores with HDDL plugin

[kumar__sumit]

Hi,

I am trying to run yolov3 model on VPU cores with the help of HDDL plugin. I am using object_detection_demo_yolov3_async application but i am getting 1.5-2 FPS which is very bad. As per the device_snapshot and utilization log, vpu cores are not being utilized 100%. I have attached this log for reference.

I also tried with the benchmark application with yolov3 model and i got VPU cores utilization as 100% and throughput of 38 FPS.

Can you please tell me how benchmark application is using all vpu cores completely and giving far better throughput compared to object_detection_demo_yolov3_async and other applications?

And Do i need to modify anything in source code of these application in order to use vpu cores completely?

[JesusE_Intel]

Hi Sumit,

Let me take a look at the YOLOv3 model with HDDL and get back to you. I should have an update in a couple of days.

Regards,

Jesus

[JesusE_Intel]

Hi Sumit,

I apologize for the delay, the object_detection_demo_yolov3_async was not developed for HDDL and it only uses 2 infer requests (hard-coded). You should be able to modify the source code of the demo to use additional infer requests.

The following link contains programming guides for HDDL devices.

https://docs.openvinotoolkit.org/latest/_docs_install_guides_movidius_programming_guide.html

Regards,

Jesus