Normalization to [-1, 1] for DeepLab

[netw0rkf10w]

Hi @nyoki-mtl,

I would like to ask a quick question please.

I see in the code that, for DeepLab, you normalize the input image to the range [-1, 1] like this:

image = (image/255)*2 - 1

Could you please tell me how you came up with this? I did some experiments and this normalization indeed yields the best performance. I checked the official TensorFlow implementation but I couldn't find where they used this normalization.

Thank you in advance for your response!