Hi, Conor!

Lets consider SLIC-algorithm described on the page you referenced. As it is noted in docs, the quickshift algo is closely related to SLIC one.

So, where 5D in these algorithms comes from? When processing a regular image, it is represented as WxHx3 multidimensional array, where third dimension respects to  color. If M is such multidimensional array, then M[i,j,0] will describe intensity of RED color at (i,j) pixel; M[i,j,1] will describe intensity of GREEN color at (i,j) pixel, and, finally, M[i,j,2] will describe intensity of BLUE color at (i,j) pixel;  

Now, we can consider an image as a heap of pixels; each pixel is characterized by its position (i,j) and color intensities (R, G, B); So, we can think a pixel as an vector in the 5D space (i, j, R, G, B);  SLIC is just k-means clustering approach applied to the heap of pixels represented as vectors in 5D space.

When you are working with multiband raster images, e.g. 10 band, resulting space dimension will be 12: 2 + 10 ( for normal, 3-band, RGB images it was 2+3 = 5D, this is noted in the sklearn docs)

Unfortunately, quickshift algorithm cann't handle multiband raster images (only 3-band, RGB-ones), but SLIC can!

# Consider the following Python code with pregenerated sample image

import numpy as np

from skimage.segmentation import slic, quickshift

tenband_image = np.random.rand(100,200,10) # a 100x200 10-band sample image

slic(tenband_image) # will return an array of shape 100x200 of labels for each pixel

quickshift(tenband_image) # will raise an error... 

quickshift(np.random.rand(100,200,3)) # pass without errors.. 

Hope that helps!

Thanks Dmitry - I have been testing the Quickshift by setting the convert2Lab bool to false, which allows you to use the multiband images without error but I was not sure how many bands it was using.

So to clarify - 

2D singlechannel = ?

2D multichannel = ?

3D = greyscale? (x,y, 1 band)

4D = ?

5D = RGB (x,y,r,g,b)

6D+ = x,y,r,g,b,NIR, etc

Hi, Conor!

I think, there is some misunderstanding due to mixing "a real life" dimension, and "internal" dimension occurring when the algorithm operates.

2D singlechannel = e.g. a normal grayscale image; (Real life dimension is 2D, but internally each pixel could be think as a vector in 3D space (x,y, gray), where x,y -- pixel position; gray -- gray color intensity)

2D multichannel = e.g. a normal RGB-image (R, G, B -- channels), CMYK-images (in this case, internally each pixel could be treated as a vector in N+2 space, where N -- the number of bands/channels;

3D singlechannel = e.g. a point could -- obtained with help of LiDAR scanning equipment; Each pixel could have its own intensity; In this case, segmentation algorithm will operate  in x,y,z + some-feature-intensity = 4D internal space;

3D multichannel = e.g. some LiDAR devices equiped with camera allows to get a color for each point in the point cloud; If the color is encoded with R,G,B, we lead to 6D internal dimension (x,y,z, R,G,B);

4D -- no real life examples, may be a point cloud changing in time...

5D, 6D+ -- it is hard to come up with..

==========

When we are working with 2D multichannel images, each pixel has a set of features: e.g. R,G,B -- intensities, NIR, FIR; In this case, internally each pixel could be treated as a vector (x, y, R,G,B, NIR, FIR), i.e. in 7D space (not a real life space, but abstract).