[feature] tf.contrib.image.invert about addons HOT 13 CLOSED

Could you give an example of how this would be used? Also, how would you imagine implementing this. It seems to be it would be difficult, because you'd have to introspect a general tensor to make it work like you suggest.

from addons.

Not sure what you mean about introspecting tensor. I would think this would just convert the transform, represented as a tensor of shape (num_images, 8), to the standard matrix representation of an affine transform: num_images x 3 x 3, and then doing a standard matrix inverse.

For some more info:

• http://scikit-image.org/docs/dev/api/skimage.transform.html#affinetransform
• https://en.wikipedia.org/wiki/Transformation_matrix#Examples_in_2D_computer_graphics

This would be used in the case where you want to take an input image, transform it, run that transformed image through a network, get a result (may be coordinates or an output image) and then transform the results back to the original space.

from addons.

I don't think this is feasible in general. I'll close this issue. If someone can come up with a workable design, please file a new issue.

from addons.

Maybe I confused matters. This should just be a matter of reshaping the 8 values to 9 and then calling tf.matrix_inverse.

from addons.

In fact that reshaping method exists: _flat_transforms_to_matrices it just isn't exported.

from addons.

This also seems feasible to me.

from addons.

@ringw what do you think about exposing _flat_transforms_to_matrices ?

from addons.

That's fine with me, if we also expose _transform_matrices_to_flat for consistency.

from addons.

This sounds fine. @cancan101, I was confused that you meant "Transforms" in the general sense ... where contrast and brightness are a type of "transform" and you wanted to be able to create an op that undoes things. Would you be willing to submit a pull request @cancan101.

from addons.

I've run into numerical instabilities inverting transformation matrices before, though for data augmentation it should suffice. Usually, one can construct the inverses using the inverse primitive affine transformations (shear, rotation etc.) in reverse order to achieve better numerical stability. Perhaps a function would be helpful that either constructs the mapping or the inverse like this tf.image.affine_transform(shear_x, shear_y, rotation, scale, translation_before, translation_after, inverse=False). Not entirely sure how useful that is. Just an idea.

from addons.

from addons.

Though supporting tf.image inversions is outside of the scope of TensorFlow Core, it might be a welcome contribution to image processing in TensorFlow Addons.

Transferring this issue to the tensorflow/addons repo; @seanpmorgan for visibility.

from addons.

My initial thought is that this could be difficult to maintain if it applied to all of the transforms in tfa.image. Seeing as the last comments were in 2017 I'm going to close this issue but happy to re-open and discuss if there is still a desire for this.

from addons.