def test_2d(timing=False, verbose=False, print_reg_params=False):
if timing is True:
tic = time.time()
try:
X = np.loadtxt(os.path.join(dir_path, "..", "data", "fish_target.txt"))
except OSError:
raise Exception("Error finding data!")
try:
Y = np.loadtxt(os.path.join(dir_path, "..", "data", "fish_source.txt"))
except OSError:
raise Exception("Error finding data!")
reg = deformable_registration(
**{
"X": X,
"Y": Y,
"verbose": verbose,
"print_reg_params": print_reg_params,
"max_iterations": 500,
})
TY, _ = reg.register()
assert_array_almost_equal(X, TY, decimal=1)
if timing is True:
toc = time.time()
print("Test 2D Affine took on fish took: {}".format(toc - tic))
def test_3d(timing=False,
print_errors=False,
verbose=False,
print_reg_params=False,
one_percent_error=0.5,
five_percent_error=0.4,
ten_percent_error=0.35):
if timing is True:
tic = time.time()
try:
X = np.loadtxt('../data/surface_points_bone_deformable_target.npy')
except OSError:
X = np.loadtxt('data/surface_points_bone_deformable_target.npy')
# Below are points from a completely different knee that were already rigidly registered to X
# If there isnt something to make them "somewhat" close to one another, then the registration fails.
# Therefore, this first step was performed to improve testing.
try:
Y = np.loadtxt('../data/surface_points_bone_1_5k_points.npy')
except OSError:
Y = np.loadtxt('data/surface_points_bone_1_5k_points.npy')
# These will not perfectly align and they will not even be "done" when we get to iteration 100.
# But this is a good starting point test.
reg = deformable_registration(
**{
'X': X,
'Y': Y,
'max_iterations': 100,
'alpha': 0.1,
'beta': 3,
'verbose': verbose,
'print_reg_params': print_reg_params
})
TY, _ = reg.register()
differences = X[:, None, :] - TY[None, :, :]
distances = np.sqrt(np.sum(differences**2, axis=2))
min_x_dist_per_ty_point = np.min(distances, axis=0)
sorted_distances = np.sort(min_x_dist_per_ty_point)
worst_one_percent_error = sorted_distances[int(
len(min_x_dist_per_ty_point) * 0.99)]
worst_five_percent_error = sorted_distances[int(
len(min_x_dist_per_ty_point) * 0.95)]
worst_ten_percent_error = sorted_distances[int(
len(min_x_dist_per_ty_point) * 0.90)]
if print_errors is True:
print('Worst 1% error: {}'.format(worst_one_percent_error))
print('Worst 5% error: {}'.format(worst_five_percent_error))
print('Worst 10% error: {}'.format(worst_ten_percent_error))
assert worst_one_percent_error < one_percent_error
assert worst_five_percent_error < five_percent_error
assert worst_ten_percent_error < ten_percent_error
if timing is True:
toc = time.time()
print('Test 3D Deformable on knee with 5k points took: {}'.format(toc -
tic))
def register_target_to_source(self, reg_type='deformable'):
if reg_type == 'deformable':
reg = cycpd.deformable_registration(
**{
'X':
self.source_spectral_coords[
self.graph_source.
get_list_rand_idxs(self.n_coords_spectral_registration
), :],
'Y':
self.target_spectral_coords[
self.graph_target.
get_list_rand_idxs(self.n_coords_spectral_registration
), :],
'num_eig':
self.non_rigid_n_eigens,
'max_iterations':
self.non_rigid_max_iterations,
'tolerance':
self.non_rigid_tolerance,
'alpha':
self.non_rigid_alpha,
'beta':
self.non_rigid_beta
})
_, self.non_rigid_params = reg.register()
elif reg_type == 'affine':
# Using affine instead of truly rigid, because rigid doesnt accept >3 dimensions at moment.
reg = cycpd.affine_registration(
**{
'X':
self.source_spectral_coords[
self.graph_source.
get_list_rand_idxs(self.n_coords_spectral_registration
), :],
'Y':
self.target_spectral_coords[
self.graph_target.
get_list_rand_idxs(self.n_coords_spectral_registration
), :],
'max_iterations':
self.rigid_reg_max_iterations,
'tolerance':
self.rigid_tolerance
})
_, self.rigid_params = reg.register()
# Apply transform to all points (ensures all points are transformed even if not all used for registration).
self.target_spectral_coords = reg.transform_point_cloud(
self.target_spectral_coords)
def main():
X = np.loadtxt('../data/surface_points_bone_deformable_target.npy')
# Below are points from a completely different knee that were already rigidly registered to X
# If there isnt something to make them "somewhat" close to one another, then the registration fails.
# Therefore, this first step was performed to improve testing.
Y = np.loadtxt('../data/surface_points_bone_1_5k_points.npy')
# These will not perfectly align and they will not even be "done" when we get to iteration 100.
# But this is a good starting point test and shows the movement of one of the meshes over time as it tries to align
# with the other mesh.
fig = plt.figure(figsize=plt.figaspect(0.5))
ax1 = fig.add_subplot(121, projection='3d')
ax2 = fig.add_subplot(122, projection='3d')
ax = [ax1, ax2]
callback = partial(visualize, ax=ax, fig=fig)
reg = deformable_registration(**{
'X': X,
'Y': Y,
'max_iterations': 100,
'alpha': 0.1,
'beta': 3
})
TY, _ = reg.register(callback)
plt.show()
differences = X[:, None, :] - TY[None, :, :]
distances = np.sqrt(np.sum(differences**2, axis=2))
min_x_dist_per_ty_point = np.min(distances, axis=0)
sorted_distances = np.sort(min_x_dist_per_ty_point)
worst_one_percent_error = sorted_distances[int(
len(min_x_dist_per_ty_point) * 0.99)]
worst_five_percent_error = sorted_distances[int(
len(min_x_dist_per_ty_point) * 0.95)]
worst_ten_percent_error = sorted_distances[int(
len(min_x_dist_per_ty_point) * 0.90)]
plt.figure()
plt.hist(sorted_distances)
plt.show()
print('Bottom 1% error: {}'.format(worst_one_percent_error))
print('Bottom 5% error: {}'.format(worst_five_percent_error))
print('Bottom 10% error: {}'.format(worst_ten_percent_error))
def test_2d(timing=False, verbose=False, print_reg_params=False):
if timing is True:
tic = time.time()
try:
X = np.loadtxt('data/fish_target.txt')
except OSError:
X = np.loadtxt('../data/fish_target.txt')
try:
Y = np.loadtxt('data/fish_source.txt')
except OSError:
Y = np.loadtxt('../data/fish_source.txt')
reg = deformable_registration(**{
'X': X,
'Y': Y,
'verbose': verbose,
'print_reg_params': print_reg_params
})
TY, _ = reg.register()
assert_array_almost_equal(X, TY, decimal=1)
if timing is True:
toc = time.time()
print('Test 2D Affine took on fish took: {}'.format(toc - tic))
def test_3d(
timing=False,
print_errors=False,
verbose=False,
print_reg_params=False,
one_percent_error=0.5,
five_percent_error=0.4,
ten_percent_error=0.35,
):
if timing is True:
tic = time.time()
try:
X = np.loadtxt(
os.path.join(dir_path, "..", "data",
"surface_points_bone_deformable_target.npy"))
except OSError:
raise Exception("Error finding data!")
# Below are points from a completely different knee that were already rigidly registered to X
# If there isnt something to make them "somewhat" close to one another, then the registration fails.
# Therefore, this first step was performed to improve testing.
try:
Y = np.loadtxt(
os.path.join(dir_path, "..", "data",
"surface_points_bone_1_5k_points.npy"))
except OSError:
raise Exception("Error finding data!")
# These will not perfectly align and they will not even be "done" when we get to iteration 100.
# But this is a good starting point test.
reg = deformable_registration(
**{
"X": X,
"Y": Y,
"max_iterations": 500,
"alpha": 0.1,
"beta": 3,
"verbose": verbose,
"print_reg_params": print_reg_params,
})
TY, _ = reg.register()
differences = X[:, None, :] - TY[None, :, :]
distances = np.sqrt(np.sum(differences**2, axis=2))
min_x_dist_per_ty_point = np.min(distances, axis=0)
sorted_distances = np.sort(min_x_dist_per_ty_point)
worst_one_percent_error = sorted_distances[int(
len(min_x_dist_per_ty_point) * 0.99)]
worst_five_percent_error = sorted_distances[int(
len(min_x_dist_per_ty_point) * 0.95)]
worst_ten_percent_error = sorted_distances[int(
len(min_x_dist_per_ty_point) * 0.90)]
if print_errors is True:
print("Worst 1% error: {}".format(worst_one_percent_error))
print("Worst 5% error: {}".format(worst_five_percent_error))
print("Worst 10% error: {}".format(worst_ten_percent_error))
assert worst_one_percent_error < one_percent_error
assert worst_five_percent_error < five_percent_error
assert worst_ten_percent_error < ten_percent_error
if timing is True:
toc = time.time()
print("Test 3D Deformable on knee with 5k points took: {}".format(toc -
tic))