Coffeehouse Thread

28 posts

Forum Read Only

This forum has been made read only by the site admins. No new threads or comments can be added.

Realtime 3D scanning and reconstruction with Kinect

Back to Forum: Coffeehouse
  • User profile image
    Richard.Hein

    This is awesome. 

  • User profile image
    magicalclick

    And it is real time updated with changes = super awesome. Oh and it also has texture mapped later in the video, nice. And augmented reality, OMG, still more? And augmented reality that you can interact? This is just insane.

    Leaving WM on 5/2018 if no apps, no dedicated billboards where I drive, no Store name.
    Last modified
  • User profile image
    W3bbo

    Having personally conducted research in computer vision, this project isn't too impressive, in fact I'm surprised it took them so long.

    Y'see, scene reconstruction is one of the holy-grails of computer vision, however because the Kinect already has a depth-sensing camera it means that a lot of the difficult reverse-projection mathematics doesn't need to be done.

    I'm not putting them down: I recognise that what they've done is a lot of hard work, but I'll be truely impressed when they can do this level of real-time scene reconstruction and segmentation using only RGB cameras.

  • User profile image
    PaoloM

    ,W3bbo wrote

    Having personally conducted research in computer vision, this project isn't too impressive, in fact I'm surprised it took them so long.

    Y'see, scene reconstruction is one of the holy-grails of computer vision, however because the Kinect already has a depth-sensing camera it means that a lot of the difficult reverse-projection mathematics doesn't need to be done.

    I'm not putting them down: I recognise that what they've done is a lot of hard work, but I'll be truely impressed when they can do this level of real-time scene reconstruction and segmentation using only RGB cameras.

    We are waiting with baited breath for the revelation of your stupendous work.

  • User profile image
    cbae

    ,W3bbo wrote

    Having personally conducted research in computer vision, this project isn't too impressive, in fact I'm surprised it took them so long.

    Y'see, scene reconstruction is one of the holy-grails of computer vision, however because the Kinect already has a depth-sensing camera it means that a lot of the difficult reverse-projection mathematics doesn't need to be done.

    I watched the video, and I didn't notice any mention of how long it took to build this. Maybe they were able to slap it together in a couple of hours since the Kinect SDK is so awesome and makes everything so easy to do.

    I'm not putting them down: I recognise that what they've done is a lot of hard work, but I'll be truely impressed when they can do this level of real-time scene reconstruction and segmentation using only RGB cameras.

    I sometimes create HTML pages using only NotePad with no assistance from an IDE. Are you impressed?

  • User profile image
    kettch

    @W3bbo: Why does it have to be done with only RGB cameras? That seems like an artificial limitation. Sure doing it that way would be a way to enable scene reconstruction with commodity hardware. Doing it with a higher resolution and more portable version of Kinect would make scene reconstruction a marketable reality instead of purely an academic exercise. However, that marketable reality would ensure that there was enough funding to continue searching for other/better ways of doing it.

  • User profile image
    magicalclick

    @W3bbo:

    it gets the job done and does it well. So.... why even care about strange estimation math when all you want is the end result?

    Leaving WM on 5/2018 if no apps, no dedicated billboards where I drive, no Store name.
    Last modified
  • User profile image
    Minh

    @PaoloM, W3bbo will be impressed with my new project that reconstructs the scene in HD using nothing but 2 Barbie Video Girls and about a hundred lines of Javascript.

  • User profile image
    W3bbo

    ,PaoloM wrote

    We are waiting with baited breath for the revelation of your stupendous work.

    My field is segmentation and tracking, not scene reconstruction or structure-from-motion, though I've read a lot of literature in both areas. Besides, my work is irrelevant, I'm just saying that the work demonstrated in this video is quite conceptually simple and (should be) straight-forward to implement, but I commend them on their work and originality.

    I'm not putting it down or denigrating it, just pointing out that this isn't quite the holy-grail of scene reconstruction we're looking forbecause isn't done using RGB information: using a depth-sensing camera takes away a lot of the hard work, but DS cameras aren't suitable for every application: this approach only works in small enclosed spaces (like rooms), for example.

  • User profile image
    W3bbo

    ,cbae wrote

    I sometimes create HTML pages using only NotePad with no assistance from an IDE. Are you impressed?

    There's a huge difference between writing HTML in an IDE vs. Notepad compared to scene reconstruction with and without depth information. Your analogy is inappropriate.

  • User profile image
    Bas

    Christ, just admit that it's cool already.

  • User profile image
    Minh

    I also want to say that my vision research involves hundreds of hours on HotOrNot.com

  • User profile image
    W3bbo

    ,Bas wrote

    Christ, just admit that it's cool already.

    It's cool.

  • User profile image
    magicalclick

    I wonder if they do thin plate energy minimization or just simply construct flat polygons from the data. Probably just the later since thin plate is redundant with this many data points.

    Leaving WM on 5/2018 if no apps, no dedicated billboards where I drive, no Store name.
    Last modified
  • User profile image
    W3bbo

    ,magicalclick wrote

    I wonder if they do thin plate energy minimization

    The papers I'm seeing relating to TPS seems more concerned with modelling motion than scene reconstruction. I'm interested by your suggestion, do you have any further information?

  • User profile image
    PaoloM

    ,Minh wrote

    I also want to say that my vision research involves hundreds of hours on HotOrNot.com

    Actually, the thesis paper of one of the current Niners was exactly about determining the "nakedness" of an online picture... Smiley

  • User profile image
    W3bbo

    ,PaoloM wrote

    *snip*

    Actually, the thesis paper of one of the current Niners was exactly about determining the "nakedness" of an online picture...Smiley

    Littleguru's. His strategy was quite interesting: use face detection algorithms to find a face in a scene, and measure the colour of similar pixels in close proximity to gauge the scene's skin colour, then hunt for connected pixels elsewhere over time to find skin pixels.

    I wonder if anyone's applied Eigenfaces to genitalia detection...

  • User profile image
    Blue Ink

    ,kettch wrote

    @W3bbo: Why does it have to be done with only RGB cameras?

    [snip]

    Probably because we have overwhelming evidence that it can be done, at least using some biological substitutes of an RGB camera.

Conversation locked

This conversation has been locked by the site admins. No new comments can be made.