What exactly is he performing the PCA on? I assume the image is concatenated into a 1D vector, so what makes up the matrix that's used to generate the eigenvectors? RGB values?
I may be getting this wrong, but doesn't that slide show the PCA being done on multiple images? Each image is 321 x 261, which concats to a 83781 vector. Then they're repeating that for 32 images producing a 83781 x 32 matrix, which they then apply the PCA on to get 32 principle components.
When you have only one image, what's making up the columns of the matrix (assuming rows = 1D image vector)?
> I may be getting this wrong, but doesn't that slide show the PCA being done on multiple images? Each image is 321 x 261, which concats to a 83781 vector. Then they're repeating that for 32 images producing a 83781 x 32 matrix, which they then apply the PCA on to get 32 principle components.
Yes, you are right (except that it's 'principal'), and I didn't read my link carefully. I am a mathematician whose specialty is not a million miles away from these kind of signal-reconstruction techniques, but far enough away that I can't give any better answer from personal knowledge. Sorry!
