How to remove the noise pixels on the edge of a point cloud captured by a Kinect? RRS feed

  • General discussion

  • Dear all,

    I'm doing a project using Kinect to capture point clouds. I use function " NuiImageGetColorPixelCoordinateFrameFromDepthPixelFrameAtResolution "  to align the depth data and RGB data. But there is a big problem: There always exsist noise points whose colors are weird on the edge of a point cloud. Can anyone tell me how to fix this? Thank you very much in advance!

    Tian Yang

    The weird white points are in the red box

    Monday, May 5, 2014 1:19 PM

All replies

  • You can do some research on various filtering techniques to remove more stray depth pixels before you export your point cloud. In general, the data will have various noise levels based on the amount of stray IR in the room that will affect the quality.

    Carmine Sirignano - MSFT

    Tuesday, May 6, 2014 7:01 PM
  • Carmine,

    Thank you! Could you please tell me some useful filtering techniques?

    Thursday, May 8, 2014 11:40 AM
  • anything like an edge detection filter would work. You can evaluate color pixels in addition to depth values to ensure this is something you are considering to include.

    Additionally, you may want to try and smooth/fill the depth frame to ensure better continuity of pixels.

    Carmine Sirignano - MSFT

    Thursday, May 8, 2014 6:55 PM
  • OK! Thanks a lot!
    Wednesday, May 14, 2014 11:23 AM
  • I have a question regarding the newer versions of KinectFusion in v2.0. There is noise present around the edge of the point cloud/mesh obtained from KinectFusion. This is probably a limitation of the depth sensor. As I did some literature review this is axial and lateral noise can be modeled as given in "Modeling Kinect Sensor Noise for Improved 3D Reconstruction and Tracking" paper by Chuong V. Nguyen et. al. found here:

    This paper is somewhat old and I was wondering if these changes are adopted in the new original KinectFusion version in the SDK 2.0.

    Looking forward to a reply.

    Friday, May 22, 2015 6:23 PM