Transforms with scikit-image

Documentation: https://scikit-image.org/docs/stable/api/skimage.transform.html

Estimate a transform

Source and destination points are given as NxD arrays, with the points given in (x, y) coordinates, not in (row, column).

So for instance, we could have:

src_points = np.array([[100, 100],
                       [300, 400],
                       [500, 300]])

dst_points = np.array([[100, 200],
                       [20, 470],
                       [250, 500]])

tx = AffineTransform()
tx.estimate(src_points, dst_points)
print(tx.rotation, tx.translation, tx.scale, tx.shear)

Which will give us a rotation of 0.53 radians, a translation of x=101.25, y=95, a scale factor of sx=0.88, sy=0.91, and a shear angle of 0.38.

Transforming an image

The warp function allows us to apply the transform to an image:

im_ = warp(im, tx.inverse)

Note that we have to use tx.inverse, not tx: the transform in warp is from the destination to the source.

Transforming points

The transform object is callable, and can be used to compute transformed coordinates:

warped_points = tx(src_points)
print(warped_points)

Will give us the same coordinates as the dst_points in return.

Effect of shear

Shear is a bit weird. We can specify shear along x and y axis with a tuple, or just along x axis.

tx_shear_x = AffineTransform(shear=(math.pi/6, 0))
tx_shear_y = AffineTransform(shear=(0, math.pi/6))
tx_shear_xy = AffineTransform(shear=(math.pi/6, math.pi/6))

For x shear, the angle is the rotation of the y-axis, with 0 as “down” and going clockwise, etc. It’s weird.