Normalizing and unnorm don't work for SimilarityTransform and higher

[david-hoffman]

dphtools/dphtools/utils/registration.py

Line 431 in 84ec9f0

self.scale_x = self.scale_y = 1 / np.array((anisotropic_scale, anisotropic_scale))

Something is wrong with anisotropic scaling when rotations are involved.

[david-hoffman]

dphtools/dphtools/utils/registration.py

Lines 255 to 299 in 84ec9f0

	def norm_data(self):
	"""Normalize data to mean 0 and unit variance."""
	# calculate mean displacement
	logger.debug("Normalizing data")
	self.ty = self.Y.mean(0, keepdims=True)
	self.tx = self.X.mean(0, keepdims=True)
	logger.debug("tx = {}, ty = {}".format(self.tx, self.ty))
	# move point clouds
	self._Y_orig = self.Y
	self._X_orig = self.X
	self.Y = self.Y - self.ty
	self.X = self.X - self.tx
	# calculate scale
	self.calc_init_scale()
	logger.debug("scale_x = {}, scale_y = {}".format(self.scale_x, self.scale_y))
	# apply scale
	Sx = np.diag(self.scale_x)
	Sy = np.diag(self.scale_y)
	Sx_1 = np.diag(1 / self.scale_x)
	Sy_1 = np.diag(1 / self.scale_y)
	self.Y = self.Y @ Sy
	self.X = self.X @ Sx
	self.translation = (self.ty @ self.B.T + self.translation - self.tx) @ Sx
	logger.debug(f"B = {self.B}")
	self.B = Sx @ self.B @ Sy_1
	logger.debug(f"B = {self.B}")
	def unnorm_data(self):
	"""Undo the intial normalization."""
	logger.debug("Undoing normalization")
	logger.debug("tx = {}, ty = {}".format(self.tx, self.ty))
	logger.debug("scale_x = {}, scale_y = {}".format(self.scale_x, self.scale_y))
	Sx = np.diag(self.scale_x)
	Sy = np.diag(self.scale_y)
	Sx_1 = np.diag(1 / self.scale_x)
	Sy_1 = np.diag(1 / self.scale_y)
	# the scale matrices are diagonal so S.T == S
	self.Y = self.Y @ Sy_1 + self.ty
	self.X = self.X @ Sx_1 + self.tx
	assert np.allclose(self.Y, self._Y_orig), "Failed to revert"
	assert np.allclose(self.X, self._X_orig), "Failed to revert"
	# B doesn't need to be transposed and
	self.B = Sx_1 @ self.B @ Sy
	self.translation = -self.ty @ self.B.T + self.translation @ Sx_1 + self.tx

These lines seem at fault but not sure why. They seem to work for Rigid and translation.