Hi Serge
Hhere you go (attached).
Notes:
- The transform is very sensitive *outside of any coordinates specified
in the source coordinates*, so make sure it covers your whole image
- The process I follow here is as follows: estimate the transform,
calculate where the corners of the image will go, then estimate a new
transform that includes a shift that positions the result on a new,
larger output image
- Estimate `affine` (forward) and warp on `affine.inverse` (because warp
expects an inverse transformation to be provided)
I hope that helps!
Stéfan
On Thu, Apr 6, 2017, at 15:50, Stefan van der Walt wrote:
> On Thu, Apr 6, 2017, at 09:57, Stefan van der Walt wrote:
>> On Wed, Apr 5, 2017, at 19:02, Serge Shakhov wrote:
>>> Just want to clarify. Am I right that your solution isn't
>>> applicable to PiecewiseAffineTransform at all?
>>> Do you think it may be possible to find a workaround?
>>
>> Sorry, I should have been clearer. With "defined" I meant
>> implemented.
>
> Upon closer inspection, it turns out that the PiecewiseAffineTransform
> *does* have an inverse implemented. It was the "+" operator causing
> the problem.
>
> Still, I have found cases now where I get invalid answers for valid-
> seeming sets of coordinates. I'm investigating further, but I think I
> have a solution to your problem.
>
> Stéfan
>
> _________________________________________________
> scikit-image mailing list
> [email protected]
> https://mail.python.org/mailman/listinfo/scikit-image
import skimage.transform
import matplotlib.pyplot as plt
from skimage import data
import numpy as np
src_im = data.checkerboard()
src = np.array([[(0,0),(200,0),(210,0)],
[(0,210),(200,210),(210,210)]])
dst = np.array([[(0,0),(250,-20),(230,-5)],
[(30,150),(240,170),(170,170)]])
src = src.reshape((-1,2), order='F')
dst = dst.reshape((-1,2), order='F')
# Estimate transformation
affine = skimage.transform.PiecewiseAffineTransform()
affine.estimate(src, dst)
# Shape of middle image, our registration target
r, c = src_im.shape
# Note that transformations take coordinates in (x, y) format,
# not (row, column), in order to be consistent with most literature
corners = np.array([[0, 0],
[0, r],
[c, 0],
[c, r]])
# Warp the image corners to their new positions
warped_corners = affine(corners)
# The overall output shape will be max - min
corner_min = np.min(warped_corners, axis=0)
corner_max = np.max(warped_corners, axis=0)
output_shape = (corner_max - corner_min)
# Ensure integer shape with np.ceil and dtype conversion
output_shape = np.ceil(output_shape[::-1]).astype(int)
# Re-estimate transform with offset
shifted_transform = skimage.transform.PiecewiseAffineTransform()
shifted_transform.estimate(src, dst - corner_min)
dst_arr = skimage.transform.warp(src_im, shifted_transform.inverse,
output_shape=output_shape,
order=1, mode='constant', cval=0,
clip=False, preserve_range=False)
f, (ax0, ax1) = plt.subplots(1, 2)
ax0.imshow(src_im, cmap='gray')
ax1.imshow(dst_arr, cmap='gray')
plt.show()
_______________________________________________
scikit-image mailing list
[email protected]
https://mail.python.org/mailman/listinfo/scikit-image
