what data type can transforms apply to? cv2 image? numpy array? other?

[xiaoyuan0203]

what data type can transforms apply to? cv2 image? numpy array? other?

[Dipet]

Numpy array in HWC format

[xiaoyuan0203]

Dose a certain transform will indicate what type of image(grey, RGB, multichannel) it can apply to? and where?

[Dipet]

It looks like we really have a problem with this and we need to add this information to the docs.
Now, looks like the only way to be certain that transforms work with any count of channels - just check single transform.
All spatial-level transforms must work with any number of channels.
Main problems with transforms, that works on color: color shifts, hue, saturation, weather augs, etc.

[xiaoyuan0203]

I have an idea. We can add tags for each transform indicating the type of this transform. The type of transform can be:

• pixel-level: not change the position information of the image
• spatial-level: change the position information of the image
• channel-independent: apply to each channel independently(so it can apply to any number of channels of cource), such as translate, rotate...
• channel-dependent: such as color shift

[xiaoyuan0203]

And make a function to filter certein types of transforms as a search space so that kinds of composed augmentation methods such as augmix can make a transform-chain from certain types of tranforms. Augmix and other composed augmentation methods can integrated to this repo with the existing transforms

[xiaoyuan0203]

These methods don't make specific transform but offer a composition policy to compose the existing transforms. The simplest composition policy is: Compose all the transforms offered sequentially. Augmix is doing such a task. Other policy can be openly explored.