Are there any computer vision algorithms specifically targeting depth images?

Are there any computer vision algorithms specifically targeting depth images?

• I've been looking into marker detection algorithms to use with a kinect based application, and the majority of the work I've been able to find is obviously focused on feature detection in 'normal' images. However, the kinect hardware provides (essentially, once you've adjusted) an 11-bit depth value per pixel. This depth image also has various visual artifacts in it from the shadows cast around the edges of objects (see for example the strong black border in this video http://www.youtube.com/watch?v=-q8rRk8Iqww&feature=related). While some traditional machine vision techniques (eg. edge detection) work well with this, others don't, and it seems like there's little information on the net discussing this. As a simple example, using the depth value makes it trivial to detect the orientation of a marker block once you've located it. So, has anyone seen any discussions / papers / etc. that cover processing a depth image for feature detection? Can anyone recommend a good algorithm for detecting "depth" markers (effectively origami blocks instead of printed b/w markers)? What I've done so far has been adhoc experimentation using opencv to process the images, but that's no where near stable or fast enough. If you link to a commercial machine vision product without some kind of trial, please mention in your answer why you think it's appropriate.

• Answer:

  My favourite 2.5D/3D feature descriptor for registration and recognition is the http://www.ri.cmu.edu/pub_files/pub2/johnson_andrew_1999_1/johnson_andrew_1999_1.pdf (original paper + more details in http://www.ri.cmu.edu/publication_view.html?pub_id=559 and http://www.cs.cmu.edu/~vmr/software/meshtoolbox/executables.html available from CMU). Other recent advances (all searchable on-line for suitable algorithims) include: 3D-Sift, Fast Point Feature Histogram, Normal Aligned Radial Features (NARF), Depth Kernel Descriptors. Older methods simply used surface properties such as curvature and edges to identify region patches. Which is best? Depends on what you want to find, viewpoint invariance, additional clutter, etc.