Appearance-Space Texture Synthesis (PCA)

[filousnt 573]

Questions on "Appearance-Space Texture Synthesis", http://research.microsoft.com/~hoppe/apptexsyn.pdf. I am trying to make the transformed exemplar (8D) with PCA, from the 5x5x3 (75D) original neighborhood colors. But i must have missed something because my appearance-space texture is not the same as their. A resume of what i do : for each pixel of original image(Exemplar) gather the 5x5 pixels neighborhood ( x3, RGB space ) apply a gaussian weight store the neighborhood on a 1x75D Vector [ r1, g1, b1, r2, g2, b2, ... ] find 75x75 covariance matrix = transpose( Vector ) * Vector get sorted eigenvalues & eigenvectors ( QR method ) select 8 eigenvectors from the 8 highest eigenvalues projected data = Vector * 8eigenvectors end for normalize each dimension of the 8D projected data and store it.