why we need to do feature subtract mean value in sphereface/test/code/evaluation.m file?

[Jubird915]

hi all,
in sphereface/test/code/evaluation.m we do feature subtract mean value and then divide the L2_norm value for each feature.

% get normalized feature
mu        = mean([featureLs(:, valFold), featureRs(:, valFold)], 2);
featureLs = bsxfun(@minus, featureLs, mu);
featureRs = bsxfun(@minus, featureRs, mu);
featureLs = bsxfun(@rdivide, featureLs, sqrt(sum(featureLs.^2)));
featureRs = bsxfun(@rdivide, featureRs, sqrt(sum(featureRs.^2)));

I understood divide L2_norm value to embed feature to a standard hypersphere, but why we do feature subtract mean value first?

in my opinion, feature subtract mean value will impact the cosine similarity.
for example, vector [1, 1, 4] subtract mean value equal to [-1, -1, 2], this vector goes to an other quadrant, so it obviously impact the angle between two vectors

@wy1iu , thank you and your team for your great jobs