Performance comparison between Movidius USB stick 1 and Movidius USB stick 2

[GT0001]

Hi, I purchased the movidius 2 USB stick with Myriad X and i had a performance comparison with movidius USB stick 1. By testing the person-vehicle-bike-detection-crossroad-0078 network on openvino for vehicle detection (http://docs.openvinotoolkit.org/latest/_person_vehicle_bike_detection_crossroad_0078_description_person_vehicle_bike_detection_crossroad_0078.html) , downloading .xml and bin file,

I noticed that, on my custom board, movidius 2 is faster than movidius 1 of more or less 100 ms.

On PC, i don't have this kind of problem and the difference between Mov 1 and Mov 2 is around 200 ms.

Here are some comparison data in milliseconds about inference time, using person_vehicle_bike_detection_crossroad_0078 on my custom board .

MOV 1 MOV 2

427.6 392.91

423.64 392.84

429.83 393.17

431.95 392.25

429.23 392.48

430.5 392.25

428.98 393.474

Why is the difference so low? Is the CPU also used to calculate the inference or is it entirely managed by the movidius stick?

Thank you

[Aroop_B_Intel]

Hi GT0001,

I am going to try an reproduce your results on my systems. Aside from the hardware differences, are all the software components the same? (OpenVINO toolkit, OS)

Are you testing the pretrained model with the crossroad_camera_demo?

Regards,

Aroop

[GT0001]

Hi @Aroop_at_Intel​ 

All the components are the same. It's the same custom board.

I'm testing the pretrained model in a git hub project called car_detection_tutorial by intel-iot-devkit.

Here's the link https://github.com/intel-iot-devkit/computer-vision-inference-dev-kit-tutorials/tree/openvino_toolkit_2019_r1_0/car_detection_tutorial

Regards,

Giovanni

[Aroop_B_Intel]

Hi GT0001,

Can you try to use the benchmark app to run your model and compare the results of the two sticks? The benchmark application should be in the directory that you built your samples in.

Heres a link to the OpenVINO benchmark documentation: https://docs.openvinotoolkit.org/latest/_inference_engine_samples_benchmark_app_README.html

Regards,

Aroop