Ability to Apply Crop Augmentation on selected classes

[m-ali-awan]

Feature description

Recently, I came across the problem of under-represented classes in an instance segmentation dataset. And it posed such problem, eg if I am labeling house objects, and stairs-bannisters , wall-painting are classes. Now, if I simply increase images, bannisters also increase exponentially,
But on the otherhand, if I can augment the dataset with the flexibility of Cropping only around the wall-painting class, I can reduce this data-imbalance by some extent..

Possible implementation

I don't have complete solution, but I tried a bit. I think, as we can pass class-name or class- id in the boxes once we pass it to transform , somehow if we can limit for a particular-transform(possibly a new one), class-conditional-cropping, and it only randomly select from selected classes, and it crops around those objects...