On a Macbook Pro, I was getting >25Hz frame rate for the depth camera. Yes you get 640×480 depth pixels per capture (actually 632×480 because the right-most 8bits are black.. see this thread for details) but keep in mind that this is not the true spatial resolution because there are “holes” in the cloud. For example, if you hold a thin object like a pencil in front of the Kinect, it will often not show up. So the resolution is probably closer to 320×240 but that is just a rough guess. Good luck!

ps- Unfortunately I had to lock own my older blog post due to excessive spam comments

thank you for reply… ahm, possibly detect “nothing” by comparing pixelsneighborhood by rgb intensity / difference. (Like a Threshold). So if the rgb-noise amplitude (maybe seperate channels) reaches an threshholdvalue, there is “something” or “nothing” if less. I got no Kinect now… played aroud with some Photogrammetrythings… but i would really like to get some more Kinect Data to download…. if you would contact me by eMail, i will send you a ftp access, or you could upload by Rapidshare or something…to save space on your website… i would like some more seconds of complete data so i am possibly able to interpolate… and mess around with it… Maybe, you would place a larger download for me…. would be very nice of you …. Thank you… ChrisS

I am fond of Nicolas’ work. However his Kinect calibration (in which recommends taking 30 image samples for averaging the errors) is intended to align the KInect RGB camera with the depth camera so that the corresponding pixels from the RGB and depth map images reference the same part of the scene. I agree with your idea that the pointcloud noise can be averaged out if you take successive shots from slightly different perspectives. But before you can do this, you will first need to align the points from the different clouds because there will be a shift in pixel locations from the different shots. I did not want to do that because it would require more computations.. I was hoping that simple averaging would do the trick but it does not for the reason I explain below.

A key to solving this problem is to understand the source of the noise error. Is is not a random ranging (distance measurement) error as I had originally assumed. Instead, the error is due to how the Kinect computes the pixel “depths” (Most people use the word “depth” loosely. Technically, the Kinect is computing a disparity value or offset between where the IR dots are vs. they are supposed to be based on some internal reference model). When the Kinect computes the disparity value for a pixel, it actually forces the value to lie on some imaginary plane… see monika’s January 16 comment (http://borglabs.com/blog/create-point-clouds-from-kinect#comment) on my other post for details.

re: your idea of using the RGB input, how would you determine that there is “nothing appearing in rgb visual data”? Even a plain, flat surface region in the RGB image would still have variations in intensity. perhaps you can try to implementing it in code (English would not matter too much I can give you some sample data if you still don’t have a Kinect.

Ahm, you got Depth information AND RGB visual Input…
So, if the Depth (3D) Information is noisy (maybe) and theres nothing appearing in rgb visual data, 3d values could be smoothed…. so you got planes where nothing visual apears, and if there is something visual, maybe an edge… 3d results were used. Sorry for my bad english… its not my mother language… hope i am understandble…. meant as something of visual interpolation with 3d Noise values…. if x1,y1(rg B)-x2,y2(rg B)<Threshold then goto 3dsmooth…. maybe you got my idea…

Any ideas or does anyone else have a similar problem? It doesn’t throw any errors and builds all the other OpenKinect examples.

I’m very impressed by your work and I would love to see it in action.