PDF version of this entire document
As visualising the eigenvectors should require a reshape back into image form it ought not be hard to demonstrate some results more visually. What is being passed to PCA is a vectorisation of image data, where the density of the sampled points is a variable recently added to the GUI sliders (MATLAB runs out of RAM if everything is sampled) and for the time being a compromise is also made, where NaN values are cast 0 (ideally should be the result of interpolation). It's not too clear how to model/treat those because they cannot be skipped; the reshape() function will expect some sort of value in there and not all faces are alike in the sense that they cannot overlap one another perfectly.