Dali: The formula of BrightnessContrast operator

[simonJJJ]

I noticed that the formula of BrightnessContrast operator on dali is

out = brightness_shift * output_range + brightness * (grey + contrast * (in - grey))

I can't figure out why to use this one but not the regular one:

out = brightness_shift + contrast * in

Are there some optimization techniques?

[mzient]

Hello, @simonJJJ
DALI aims to be a replacement for multiple vision processing frameworks - and they tend to follow different definitions.
Gluon CV (a computer vision library for MXNet) has additive brightness and multiplicative contrast.
Torchvision has black-centered multiplicative brighness and gray-centered multiplicative contrast.
To emulate both, DALI has an operator with both multiplicative and additive brightness.

The actual operation that's performed on the pixels is simply out = fma(x, alpha, beta)
where:

x - input pixel intensity
alpha = brightness * contrast
beta = brightness_shift * output_range + brightness * (1 - contrast) * grey

[simonJJJ]

Hi, @mzient
Another question is why you use the fixed value of grey corresponding to the specific type, while torchvision use a rgb_to_grayscale function. Is that for faster speed?
Although two different methods of grey calculation for data augmentation won't make a much difference.

[mzient]

@simonJJJ To be perfectly honest, I didn't know that their meaning of "solid grey" is "mean brightness of the image" - and even if we knew that, it's unlikely we'd go with this implementations, since it requires a reduction (calculation of mean) which is quite expensive. Also, it's not suitable for videos, because it would result in shifting of the grey level from frame to frame