Do I need to use StandardScaler after Hog feature extraction scikit-image?


Code Now

I have extracted Hog features from the Fashion-MNIST and GTSRB and want to do PCA before training SVMs. The feature vectors only contain values between 0 and 1, therefore I'm not sure if I really should use the StandardScaler before doing PCA? I have the impression that the feature vectors are already normalized.

I tested two variants for each of the two datasets. Once I used the StandardScaler before performing PCA and once I did not.

from skimage import feature

hog_feat = hog(image, orientations=9, pixels_per_cell=(8,8), cells_perBlock=(3,3), block_norm='L2_Hys', transform_sqrt=True)

The accuracy of the trained SVM is a little bit better without using StandardScaler. I tend not to use the StandardScaler, but I am very unsure as it is recommended to convert mean=0 and standard deviation=1. Would not the StandardScaler be more appropriate if I wanted to combine different features vectors with different scales and map them to a consistent scale?

Continue reading...