I'm working on the following classification task: Take two patches of an image and compare for illumination invariant texture similarity. There are a couple of features which could be used to compare two image patches for texture similarity. Candidates are Local Binary Patterns (LBP) or Histograms of Oriented Gradients (HOG). Both generate a feature vector. What I'm looking for is a good metric which can be used to compare two feature vectors for texture similarity as described above.
Asked
Active
Viewed 227 times
0
-
1L2, cosine, HISTCMP_HELLINGER – berak Feb 11 '15 at 09:33
-
training a 2-class SVM on distances(same/not-same) will work even better, also funny kernels to try, like intersection or kmod – berak Feb 11 '15 at 09:45
-
Illumination invariant screams for HSV color space; just drop the V :) – Rob Audenaerde Feb 11 '15 at 11:48
-
EMD is quite robust for histogram comparison – Rosa Gronchi Feb 11 '15 at 15:06