Allow to set a custom mask interpolation method

[creafz]

This is POC to fix an issue in #850. It supports two styles for setting mask interpolation.

Option 1. Set a global value for mask interpolation through Compose:

t = A.Compose([
    A.Resize(128, 128),
], mask_interpolation=cv2.INTER_NEAREST)

Option 2. Override a value for mask interpolation for a single transform:

t = A.Compose([
    A.Resize(128, 128).set_mask_interpolation(cv2.INTER_NEAREST_EXACT),
])

[creafz]

Or as an alternative for Option 2, maybe it is better to an explicit argument for all DualTransform subclasses, such as

t = A.Compose([
    A.Resize(128, 128, mask_interpolation=cv2.INTER_NEAREST_EXACT)
])

What do you think?

[Dipet]

Maybe better to create an Enum as class field for this flag?
With Enum error message would be much more clearly.

[Dipet]

Strange condition, I am not sure that it will be work correctly, because mask_interpolation has no default value.

[Dipet]

Or as an alternative for Option 2, maybe it is better to an explicit argument for all DualTransform subclasses, such as

t = A.Compose([
    A.Resize(128, 128, mask_interpolation=cv2.INTER_NEAREST_EXACT)
])

What do you think?

I think approach better, but it is need to change __init__ method for all transforms. If we will do this, we must to add **kwargs as an argument for all transforms to have possibility to add new features simpler

[LucaBonfiglioli]

please

[ternaus]

@LucaBonfiglioli Will look into this.

[zakajd]

Heavy +1 for this PR. It's been 3 years since discussion and PR looks almost done. What's needed to finalise it? @ternaus @Dipet

The problems with resizing are very difficult to debug and lead to worse model performance. I had bad experience using Albumentations for segmentations and super resolutions tasks due to that and avoid using most of "geometrical" transformations for the same reason.

[ternaus]

@zakajd

We can do it.

As I understand, you are "Power User", meaning use the library for your task + use advanced features.

Sent you invitation to connect on LInkedIn, let's have video chat, your feedback / requests will help with the prioritization.