Normalization range

[talhaanwarch]

I am confused whether albumentation normalize between 0 and 1 or between -1 and 1. I tested it is between -1 and 1, but I assume it to be between 0 and 1.
Here is code

aug= A.Compose([
            A.Resize(224, 224),
            #CenterCrop(224,224),
            A.HorizontalFlip(p=0.5),
            A.VerticalFlip(p=0.5),
            A.ShiftScaleRotate(0.05,0.05,5), 
            A.Cutout(p=0.5),
            A.Normalize(mean=(0.5), std=(0.5)),
            ToTensorV2(p=1.0),
        ], p=1.0)

[Dipet]

Try to use this: A.Compose([A.Normalize(mean=0.5, std=1.0)]) to normalize image to range [0, 1]

[FreeFly19]

The most common values:
A.Normalize(mean=0.0, std=1.0) -> [0, 1]
A.Normalize(mean=0.5, std=0.5) -> [-1, 1]

Just for understanding:
A.Normalize(mean=0.5, std=1.0) -> [-0.5, 0.5]
A.Normalize(mean=1.0, std=1.0) -> [-1, 0]

Here is the formula from the source code:
img = (img - mean * max_pixel_value) / (std * max_pixel_value)

[Dipet]

Yes, you're right, my fault.

[STHxiao]

Hi, why need to multiply by max_pixel_value, normally standardized formulas don't have this item, looking forward to your reply!

[Foundsheep]

It seems like, it's because

1. the range of pixels in the image could vary between [0, 1] and [0, 255] etc.
2. the mean and std value such as 0.5, 1.0 are too small if used as a absolute value with range [0, 255]
3. therefore, to use the scale of the range, we need to multiply by max_pixel_value

These are my thoughts derived from a little bit of calculations

[STHxiao]

I guess it's not like that, it seems like it's because the mean on the default ImageNet dataset is calculated by dividing by 255, so it needs to be multiplied back when normalizing. You can refer to the default parameters of this API.