Occupancy masks for geometric transforms

Consideer an equivariance loss scenario: we have a source image src, we get view1 = k1(src); view2 = k2(src). Now we’d like to put on some loss on corresponding locations in view1 and view2.

For most of geometric augmentations, not all pixel locations in view1 would have correspondences in view2, and loss must be computed only for valid locations, i.e. we should not compute loss for black pixels in rotated images in https://github.com/kornia/kornia-examples/blob/master/data_augmenation_segmentation.ipynb

In this notebook is advised to use k1(img, k1._params) for repeating the sampled transformation on another input. Is there an easy way of performing an inverse transform given _params?

I think the correct expression for mask of valid locations in view is given by mask_valid_1 = k1(k2(k2(ones_like(src), k2._params), k2._params, INVERSE = True), k1._params), but one needs an easy way of doing an inverse transform.

(related https://github.com/kornia/kornia/issues/476#issuecomment-833147695)

As a workaround, I’m not manually using return_transform, and doing warp_affine with inverse transforms

[edgarriba]

absolutely - adding a return_inverse flag could be definitely quite neat. /cc @ducha-aiki @shijianjian @lferraz thoughts ?

[edgarriba]

merged in #1013