Add support for color images

[martinberoiz]

For registration is probably better to use Y channel in YUV colorspace, or average of RGB.

Quick and dirty gist here https://gist.github.com/martinberoiz/1ab5f6ec4a81680517ad2d9bd2ab9f32

[leliel12]

The forloop can be inside a joblib parallel environment. Also another aproach is to add a `color=True` or cchek the output of the ndim function. El jue., 22 de oct. de 2020 14:25, Martin Beroiz <[email protected]> escribió:

…

For registration is probably better to use Y channel in YUV colorspace, or average of RGB. Quick and dirty gist here https://gist.github.com/martinberoiz/1ab5f6ec4a81680517ad2d9bd2ab9f32 — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <https://github.com/toros-astro/astroalign/issues/57>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AACJAASJGDRJSMOK2RN5DR3SMBTHVANCNFSM4S3P7BMA> .

[martinberoiz]

I like checking for ndim but it assumes it's a numpy array. I think this should be added to find_transform which already accepts either an iterable of (x,y) pairs or a numpy array.

The input is getting more complicated to parse, but I think it's still manageable.

This also relates to #55 but if batch_register eventually calls find_transform then we only have to modify it there once.

Regarding the joblib parallel it may be an overkill, since the loop is 3 items long.

[martinberoiz]

Following @leliel12 suggestion, I think we should follow scikit-image's convention on which axis contains color channel info.

The convention seems to be that the trailing axis is the one for channel. Images loaded with pillow (PIL) also behave this way.

The only concern is for data cubes, or collection of images in different filters, that may be created this way:

>>> data_cube = np.array([data_g, data_r, data_i, data_z])
>>> data_cube.shape
(4, 256, 256)

Then the "channel" is the leading axis. Also astropy FITS data cube seem to put the spectral index at the front as well.

A compromise could be adding an argument channel_axis=-1 that user can change to 0.
I'm reluctant because of the argument pollution but I don't see a better option.