I have a use case where an image flipped horizontally actually changes the class/label.
And allow end-user to submit some logic to that would be awesome.
Example would be a sign with an arrow… Left-sign … Would be Right-sign when flipped.
I think you can do that, it would just be without adding certain augmentations when training your model on the dataset since augmentation is a step that basically does what you’re saying would be useful without.
Edit: Ah okay so an option to void the class change despite the different position the objects might be in during augmentations. That is a good idea.