cluster RGB data

[#1]

In JMP it's possible to load pictures as tables - and Jed Campbell shared some cool tricks how to extract the data from a graph-picture:

From Picture to Data: Extracting graph data from a picture into a JMP Table for analysis 

One of the tricks: Use clustering to separate axes and grid lines from the actual data. This approach can also be used to separate data points that belong to different curves. If you know the exact number of clusters, you can use K Means Cluster, or if you want to play around, you can use Hierarchical Clustering.

I just tried it with the below data set

and was surprised by the result which I got with the default settings:

After some trial and error, I found the solution:
The range 0-1 has some meaning, by rescaling the columns individually, this information is lost.

After disabling the option:

I got a perfect separation of the data points:

For Hierarchical Clustering, one can enable 2-way-clustering:

which relies as well on data "on the same scale". For hierarchical Clustering, "single" seems to work as well:

What is your experience with clustering?

[#2]

You might also want to investigate other color spaces; transforming the RGB to HLS can be useful.

One problem with RGB is interpolating between two colors...half-way between two colors tends to be gray because the center of the RGB color cube is gray. HLS interpolates a hue axis that is independent of the gray defined by lightness and saturation. There are other color spaces; some of them are supposed to model human vision and might be used for image compression algorithms (leave out details we can't see). 

https://en.wikipedia.org/wiki/Color_space 

https://en.wikipedia.org/wiki/HSL_and_HSV  JMP uses HLS for HSL