I have a question of your paper's formulation，looking forward your reply.

[lijing-coder]

In your paper , formulation 9 has a part which is µMi, why we should add this part in this formulation? How to calculate µ?

[Fu0511]

In your paper , formulation 9 has a part which is µMi, why we should add this part in this formulation? How to calculate µ?

Sorry for the late reply, I haven't noticed the issues in this git for a long time.
It is known that, in the framework of classification, the input is usually mean centering (subtract the overall mean of the whole dataset, e.g., it is [0.485, 0.456, 0.406] for the ImageNet).
As described in our paper, µ in the formulation 9 exactly represents for the overall mean of the whole dataset. By adding µMi in the formulation 9, it can be easily verified that the overall mean of the modified images is also equal to µ. Thus, we can use the original input normalization for the modified images.
Hope that explanation helps.