superGMM results
Post date: Sep 17, 2011 6:34:57 PM
Here is the result for superGMM
feature vector to test
patch of 5 x 5 pixels
F1 = (gRGB, sLuv, gsLab, sX, sY) --> 11D
F2 = (gRGB, sLuv, gsLab, sX, sY, Texton) --> 53D!
F3 = (gRGB, sLuv, gsLab) --> 9D
PCA10_F1: 10-component PCA of [ 10-bin histogram of F1 + average F1 + median F1]
PCA10_F2: 10-component PCA of [ 10-bin histogram of F2 + average F2 + median F2]
single_F3: single point of F3 + sX, sY
avg_F3: average F3 + sX, sY
3binHist_F3: 3-bin histogram of F3 + sX, sY
The best PRI score with the best scale parameters
* We use only C from 5-8 and S from 60, 80 and 200 to test the performance.
med_F3: median F3 + sX, sY
avg_<each of F3>: average for (each of F3) + sX, sY
med_<each of F3>: median for (each of F3) + sX, sY
PCA10_10binHist_<each of F3>: 3-bin histogram of (each of F3) + sX, sY