On Mon, Apr 22, 2013 at 10:02 AM, Brickle Macho <bricklemacho@gmail.com> wrote:
I have a RGB-D image. Using only RGB I segment the image into superpixels using SLIC and Quickshift algorithms provided in scikit-image. I am trying visit each superpixel, calculate some depth features for each superpixel. Specifically I want to calculate the surface normal of the superpixel and the average angular difference with the neighbouring superpixels. Eventually I plan to combine the superpixels based on these depth features.
If you had a mask for an individual superpixel, and indices into your array, x, y, z, you can imagine finding the coordinates of all pixels under that mask with x[mask], y[mask], z[mask] The mask you typically recover from a label image, so, e.g., mask = (labels == 3). Now, the trickier problem is figuring out where, relative to other super-pixels, this one is located. For that, it may be better to represent the image as a graph, where each node represents a super-pixel, and edges represent links to other super-pixels (in fact, this is something we should implement in scikit-image to make handling labels easier). Would you be interested in collaborating on such a feature? Stéfan