Why do you normalize each img's pixel value to -1.6~1.6 when training and testing?

[yangtian62]

Hi, thanks for your inspiring work! Just one point I don't understand; why do you set
img = np.array(img, np.float32).transpose(2,0,1)/255.0 * 3.2 - 1.6
when training? Is there any advantages for this operation?

[zlckanata]

This operation is to norm the image from [0,255] to [-1.6,1.6], you can norm the image to other range(better zero centered), [-1,1] or [-1.2,1.2] are fine :)

[yangtian62]

Thanks for your reply~ Is 1.6 a particular number that works the best in your case? And why don't you apply a std to each channel? Cause people always set mean and std at the same time. Haha~

[zlckanata]

In dlinknet, each conv-layer is followed by a batchnorm(conv-bn-relu), and, in fact, there's no difference between [-1.6,1.6] and [-1,1](or even [0,255]) in this case.
However, Unet(weight initialized by pytorch default initializer) without batchnorm do have better result using [-1.6,1.6] normalization. In this case, [-1.6,1.6](compare with [-1,1]) is equal to enlarge the initial learning rate and std of the initializer.
In dlinknet, channel-wise normalization equals adding two parameters weighting the RGB. It may have better result(I'm not sure), but I don't think this is so crucial~ 2333

[yangtian62]

Alright, thanks again. Happy New Year btw~ Lol

[zlckanata]

Happy New Year~ www