search

ox-vgg / vgg_face2

646 stars 114 forks source link

Images Normalization #29

Closed wfcb-85 closed 5 years ago

wfcb-85 commented 5 years ago

Hello,
My question is mostly theoric : You mentioned in another issue that your normalization steps include only the mean substraction from each channel of each image. However, according to some people, e.g., http://josh-tobin.com/assets/pdf/troubleshooting-deep-neural-networks-01-19.pdf, not having input values between 0 and 1 could hurt performance. So my question is why are your input values much bigger than 1. Thanks in advance.

WeidiXie commented 5 years ago

Hi, That's usually not a problem in computer vision now, since you have batch norm layers.