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!

Intel Euclid

Does this device work well with Euclid to increase its object recognition ? Seems like the ultimate pair.
The pi 3 results are rather disappointing.

Comments

  • 3 Comments sorted by Votes Date Added
  • @chicagobob123 the minimum host machine requirements for NCS is an x86_64 machine running Ubuntu 16.04 with 1GB RAM and 4GB free storage. The Euclid specs are mentioned here - https://click.intel.com/intelr-euclidtm-development-kit.html. So the quick answer to your question is - yes it can work with Euclid.

    The pi results are rather disappointing

    I'm curious to know why you say so? Entire inference is offloaded to the NCS, the only thing that RPI is doing is grabbing images from it's source (camera or filesystem), sending it to the NCS and reading the results. Are you running into USB 2.0 bandwidth limitation? Or you doing other application level tasks that might be taking up your RPI CPU/memory resources?

  • I can not even get a unit yet (Mouser on back order so bad they told me they would contact me when they have them) but am drawing information from a prior post which states 0.321s to draw a single inference.
    Is that the speed with color frames at you expect to get? Do you think the USB 2.0 to be a limitation?

    Most video arrives at 29.97 NTSC and 25.0 PAL and each frame comes in at ~ .033 or 0.025 seconds
    .321s is far from real-time evaluation which makes me wonder how other systems are doing 6DoF.

  • @chicagobob123 image size and network complexity/size play a major role in inference performance. A network that uses input images of size 227x227 and has 1000 classes would be much slower than a network that uses 32x32 and has 50 class. Image size and object classes are application dependent. For ex - a security camera may not require image processing at 1080p (maybe just recording at 1080p), and wouldn't require classes like zebra, lion, aeroplane, etc. In such a case the network architect would train a model/network that has only the application specific classes with an input image size large enough to get acceptable accuracy.


    iamintel

Sign In or Register to comment.