Friday, March 11, 2011

Noise in the Kinect Depth Image

I've been looking into the noise that you get in the depth images that come from Kinect. I've found two good references so far: Kinect Z Buffer Noise and Audio Beam Steering Precision and Experiment to remove noise in Kinect depth maps. The general consensus seems to be that the error in the Kinect images cannot simply be averaged out over time, and that it has to do with some kind of quantization noise in the stereo matching algorithm. Also, most of the noise is in the center of the depth range at a few meters away. There might be some way to remove the quantization noise if it's constant with respect to the 2d image -- if it's constant with respect to the 3d space, it would be way too intensive to sample.

2 comments:

  1. This comment has been removed by the author.

    ReplyDelete
  2. This can be definitely improved...I would suggest looking into multiscale approaches such as wavelets or linear (or not) scale-space theory. Edge preserving methods would probably do the trick. An article that contains a method that I suspect might be helpful and isn't too hard to implement is this:

    www.scielo.br/pdf/cam/v24n1/08v24n1.pdf

    and another one that appears much harder to implement and uses smoothed normals is this:

    ftp://ftp.math.ucla.edu/pub/camreport/cam03-03.ps

    ReplyDelete