Alastair M. Robinson wrote:
Hi :) Klaus Karcher wrote:any hints how to calculate the median for multidimensional (tristimulus or spectral) data? I've been looking for an algorithm that doesn't tear apart the measurement dimensions, but my efforts were unsuccessful.Well I guess you'd have to digest each sample into a single-dimensional value to find the median - but in the context of GPLin, the misreads we were seeing were instantly obvious on a plot of input vs. L*, so I'd be surprised if in practice merely using the L* of Lab or Y of XYZ didn't turn out to be sufficient. Perhaps using dE from Lab 0,0,0 (or maybe even the Lab of 50% grey?) might be more robust with yellows.
I use the distance (Delta E) to the average as weighting factor in my weighted average function for R.
Maybe constructing an ellipsoid somehow around the "measurement cloud" (or it's gravity center?) might be a robust method to eliminate outliers -- but I'm afraid that's over my head.