frame

Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Sign In

Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Pose Estimation Performance

Hi - we're looking at pose estimation (in the sense of skeleton tracking) on the Movidius - excited to see that there's a pre-trained model and example code, but I can't find any sense anywhere of how fast that runs. I realise that what it's attached to will have significant bearing, as will the size of the image it's processing, but does anyone have any finger-to-the-wind idea (or better still numbers from an actual rig) of how many frames per second we could get (even if it's seconds per frame!)?

Many thanks

Mike

Comments

  • 7 Comments sorted by Votes Date Added
  • Really? Has nobody run this example on anything? I just want a sense of how fast or slow it is!

  • Hi @mikepelton

    Is there any particular example you are interested in? I can do a quick test and share the results with you.

    Regards,
    Jesus

  • Hi @Jesus_at_Intel
    I'm interesting in a pose estimation project using Movidius. I wanna ask you if you have done some work about pose estimation mission on Movidius. I‘m grateful if you share some useful models with me.
    Best wishes,
    Neil

  • Hi @neilwu

    I have not done any work with pose estimation with the Neural Compute SDK. However, I ran the example code with the pre-trained model included in the OpenVINO toolkit and I am getting about 2.8 FPS on the Intel Neural Compute Stick and about 4.5 FPS on the Intel Nerual Compute Stick 2. I am running the example code with a web cam feed.

    ./human_pose-Estimation_demo -i cam -m ~/intel/computer_vision_sdk/deployment_tools/intel_models/human-pose-estimation-0001/FP16/human-pose-estimation-0001.xml -d MYRIAD

    Running the same example code on my Ubuntu virtual machine using the CPU resulted in about 8.8 FPS.

    ./human_pose-Estimation_demo -i cam -m ~/intel/computer_vision_sdk/deployment_tools/intel_models/human-pose-estimation-0001/FP32/human-pose-estimation-0001.xml -d CPU

    Hope this helps!

    Regards,
    Jesus

  • Hi @Jesus_at_Intel
    Thanks for your advising! And I successfully ran the example human pose estimation code in the OpenVINO toolkit.
    But I found a little questions on the running speed between NCS and NCS2.
    With a local video file (means -i /PATH/TO/VIDEO/.mp4) , I am getting about 2.8 FPS on the Intel Neural Compute Stick and about 4.8 FPS on the Intel Nerual Compute Stick 2 just exactly like yours.
    BUT, with a USB cam input (means -i cam), I am getting about 1.9 FPS on the Intel Neural Compute Stick and about 2.4 FPS on the Intel Nerual Compute Stick 2. You can see the running speed improvement is little. Can you tell me about it? Thanks a lot.

    Regards,
    Neil

  • Hi @neilwu

    Could you tell me which USB camera you are using? I'm wondering if the performance is different with a higher resolution camera. The camera I used in my test had a 720p resolution.

    Update: I ran a quick test with a higher resolution (1080p) camera and saw similar FPS as my previous test. I am using OpenVINO toolkit R 5.0.1, ran tests on a Windows install of OpenVINO and an Ubuntu VM install.

    Regards,
    Jesus

  • Hi @Jesus_at_Intel

    As you say, I think the USB camera I used might be a lower resolution(a cheap camera). After I will use a higher resolution camera and test on it.
    Thanks your reply!

    Regards,
    Neil

Sign In or Register to comment.