Remove OpenCV dependency

[himbeles]

• Replaces OpenCV methods with numpy and scipy implementations
• Currently breaks the tests involving quasi-exact angle and image comparisons. Perhaps because of Nearest Neighbor interpolation in OpenCV vs. Spline Interpolation in Scipy?

[nelsond]

Unfortunately, image rotation with scipy.ndimage seems to be really slow.

import numpy as np
import timeit

from gridfit.utils import rotate as opencv_rotate
from scipy.ndimage import rotate as scipy_rotate

data = np.random.rand(1000, 1000)

Running timeit on the following code snippets

opencv_rotate(data, np.random.uniform(0, 360))

and

scipy_rotate(data, np.random.uniform(0, 360), reshape=False, order=0)

yields

1.38 ms ± 22.1 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

and

16.1 ms ± 278 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

[himbeles]

Yes, I have noticed that too :(
But to have a fair comparison, warpAffine should probably be called with the cubic interpolation flag.
Also, the interpolation used for rotation really makes a significant difference on the extracted angle (~0.3 deg).

https://docs.opencv.org/3.4/da/d54/group__imgproc__transform.html#ga5bb5a1fea74ea38e1a5445ca803ff121

[nelsond]

The difference actually seems to be negligible for cubic interpolation. I see an increase to 1.93ms (from 1.68ms), is this consistent with your observations?

It still seems that scipy.ndimage.rotate would make things ~10x slower. Considering that find_dominant_angle already takes up to a second for a reasonably sized image, this might become quite annoying depending on the use case.

Maybe making opencv an optional dependency would be a reasonable solution?