Support extended mode on projective transform

[WindQAQ]

Closed #1318. Support extend mode on projective transform. Mainly follow the scipy's implementation.

From top to bottom: original, reflect, constant, mirror, nearest, wrap. Addons is on the left, scipy is on the right.

Testing script:

import numpy as np

import tensorflow as tf
import tensorflow_addons as tfa

from scipy import ndimage, misc

image = misc.face()

tf.io.write_file("test.jpeg", tf.io.encode_jpeg(image))

for mode in ["CONSTANT", "REFLECT", "WRAP", "MIRROR", "NEAREST"]:
    t = tfa.image.rotate(image, np.pi / 4, extend=mode)
    tf.io.write_file(mode.lower() + ".jpeg", tf.io.encode_jpeg(t))
    st = ndimage.rotate(image, 45, mode=mode.lower(), reshape=False)
    tf.io.write_file("scipy-" + mode.lower() + ".jpeg", tf.io.encode_jpeg(st))

[bot-of-gabrieldemarmiesse]

@mels630

You are owner of some files modified in this pull request.
Would you kindly review the changes whenever you have the time to?
Thank you very much.

[AakashKumarNain]

@WindQAQ overall looking good but I have one query. Please check the inline comments I have made

[WindQAQ]

Just figure out that there is another PR #1719. @srilman what do you think about this :-) Thank you.

[srilman]

@WindQAQ Looks great to me! Definitely preferable to #1719. Only thing is that it would be nice if the different extended modes could be used separately (outside of projective transforms) for more flexibility.

For example, I mentioned that I liked to extend the image before performing some image ops like tf.image.random_crop. In general, extend can be used in lieu of tf.image.resize_with_pad in nearly every situation.

[WindQAQ]

@WindQAQ Looks great to me! Definitely preferable to #1719. Only thing is that it would be nice if the different extended modes could be used separately (outside of projective transforms) for more flexibility.

For example, I mentioned that I liked to extend the image before performing some image ops like tf.image.random_crop. In general, extend can be used in lieu of tf.image.resize_with_pad in nearly every situation.

We can have both of them actually if they are both useful :-)

[WindQAQ]

@bhack It seems that it is migrated to core TF in Jan. I have no idea about this. Thanks for pointing out!

cc @tensorflow/sig-addons-maintainers

tensorflow/tensorflow@2699281

[WindQAQ]

@WindQAQ I know that I am blamed as the cross-linking guy

Never mind 😄 I don't even care about it LOL. Your links are super helpful for me.

but I suppose it was part of a more general effort that we already commented at #1226 (comment) and in some follow-up messages in that thread mentioning Keras.

Umm, it seems to be the miscommunication between us and many places. A related issue has been posted #1779. Thanks again for pointing out it :-)

[bhack]

@WindQAQ Users are looking for some of these filling parameters at tensorflow/tensorflow#39177