Albumentations intensity augmentations disrupt the image

Question

I'm using a preprocessed, z-score normalized list as the source for my dataset.

Here's a collage of images augmented by Albumentations:

enter image description here

Here's my Compose:

augmentation = A.Compose([
    A.HorizontalFlip(),
    A.RandomBrightnessContrast(brightness_limit=(-0.0001, 0.0001), contrast_limit=(-0.01, 0.01)),
    A.CoarseDropout(8, 0.1, 0.1),
    A.Rotate(limit=15),
    A.Affine(shear=(-2, 2), scale=(0.95, 1.05)),
>! ToTensorV2()
])

On the 50% of the images that RandBrightnessContrast is applied even with very small parameters, the whole distribution of the image is squashed to [0, 1] (from ~-2,~2 as expected for z-score normalized images). Any way around this?

Maybe I should perform z-score normalization after these, but my original intent was to separate all deterministic steps (resize, normalize etc.) from the augmentation steps for efficiency.

Answer 1

Albumentations indeed, treats images of the type float32 as having min value of 0 and max value of 1.

If your images are in [-2, 2] range, you may rescale your image to [0, 1] with:

def convert_to_01(image):
    return (image + 2) / 4    

and later rescale back with

def convert_from_01(image):
   return (image * 4) - 2

If possible, you may even convert images to [0, 255] uint8 type, as transforms in uint8 happen much faster.

P.S. Thank you for the question, I will make it clear in the documentation of Albumentations that float32 images are expected to be in [0, 1] range.