[FEAT] Imviz: Rotate image

[pllim]

Description

User wants to rotate an image in several ways:

1. N-up E-left
2. Arbitrary angle

Out of scope: Handling distortion accurately.

Current status: @astrofrog is going to present a proof-of-concept on glue-side by 2022-07-12

cc @orifox and @kecnry

Additional context

Previous failed attempts:

• Rotate image in Imviz #1340 (too complicated but works)
• POC: canvas rotation #1384 (elegant but had performance issues)

🐱

[astrofrog]

I've been experimenting with an approach that I think could work - basically internally in glue we would have the ability to apply an arbitrary 2D affine transform to the 'viewport'. To demonstrate this, I have the following PRs:

• Experimental image rotation glue-viz/glue#2310
• Experimental image rotation glue-viz/glue-jupyter#315

If you install these two then in the Imviz example notebook you can do:

from matplotlib.transforms import Affine2D
affine_transform = Affine2D().rotate_deg(30)
imviz.app.get_viewer_by_id('imviz-0').state.affine_matrix = affine_transform

and the image will then be rotated. Selections will still work fine. In fact you should in principle be able to do for example:

affine_transform = Affine2D().rotate_deg_around(2048, 1024, 30)

which includes a translation component in the affine transform but this doesn't seem to work properly at the moment in the sense that selections are a bit offset compared to where they are drawn (but I think it should be easy to fix). Ignore for now the fact that this involves Matplotlib transforms - this is just a hack for quick development but won't be in the final code

Does this seem reasonable? Can someone try it out to let me know if it would be fast enough? @pllim if you had already made a plugin to set the rotation you could try and hook it up to this?

I decided on this approach as opposed to adding/removing virtual Data objects as that gets a lot more complicated and updating the link graph etc can be slow. If the approach here works I think it should be a lot faster.

Another caveat is that for now the mouseover coordinates are wrong but that will be an easy fix if we agree on this approach.

[pllim]

Thanks! I could try out your proof-of-concept but not this week, also depends on what will take priority next sprint. I suspect support for JWST data release would take precedent for the near future, whatever that may be.

[pllim]

See #1551

[pllim]

Done in #2179

[stscijgbot-jwql]

This issue is tracked on JIRA as JDAT-2574.

[stscijgbot-jwql]

Comment by Ori Fox on JIRA:

Here's an update. I found some code that should be able to align this and create a new fits file with the correct padding.
import matplotlib.pyplot as plt
from astropy.io import fits
from astropy.wcs import WCS
from reproject import reproject_interp
from reproject.mosaicking import find_optimal_celestial_wcs

Import Data

image = fits.open('iebc16030_drc.fits')

Find WCS where North is pointing up

wcs_in = WCS(image[1].header)
data = image[1].data
wcs_out, shape_out = find_optimal_celestial_wcs([image[1]])

Reproject image

reprojected, foot = reproject_interp((data, wcs_in), wcs_out, shape_out)

plt.imshow(reprojected, origin = 'lower', vmin = -0.038, vmax = 0.098)
plt.grid(color='white', ls=':', alpha=0.7)
plt.xlabel('RA')
plt.ylabel('DEC')
plt.show()

[stscijgbot-jwql]

Comment by Ori Fox on JIRA:

!image-2022-12-05-11-18-20-033.png!

[stscijgbot-jwql]

Comment by Pey-Lian Lim on JIRA:

Sure, you can do that outside of Jdaviz all you want. I would not recommend doing this every time someone click "change angle" inside Jdaviz. This is a very expensive operation.

[stscijgbot-jwql]

Comment by Ori Fox on JIRA:

I spoke with Tom yesterday after the demo and mentioned that before we move forward, I would like to see the following:
Reproject (pix_to_pix) applied to the entire array every time someone clicks change angle. A user will typically only want to do this once per image. Tom should test it on a 1 GB and 5 GB image and report the times.

Right now, our cube model fitting takes 1-2 minutes. So I can't imagine it taking longer than that, especially for an average size array.