def generate_random_transform(params, shape):
##
## Translation
##
translation = np.eye(params.image_dimension + 1)
for t, ax in zip(params.translation_range, params.img_spatial_indices):
translation[ax, params.image_dimension] = np.random.uniform(
-t, t) * shape[ax]
##
## Rotation
##
rotation = np.eye(params.image_dimension + 1)
if params.image_dimension == 2:
theta = (
np.pi / 180.0 *
np.random.uniform(-params.rotation_range, params.rotation_range))
rotation[:2, :2] = dv.rotation_matrix_from_angle(theta)
elif params.image_dimension == 3:
x = dv.rotation_matrix_from_angle(
np.pi / 180.0 * np.random.uniform(-params.rotation_range[0],
params.rotation_range[0]),
matrix_type="roll",
)
y = dv.rotation_matrix_from_angle(
np.pi / 180.0 * np.random.uniform(-params.rotation_range[1],
params.rotation_range[1]),
matrix_type="pitch",
)
z = dv.rotation_matrix_from_angle(
np.pi / 180.0 * np.random.uniform(-params.rotation_range[2],
params.rotation_range[2]),
matrix_type="yaw",
)
rotation[:3, :3] = np.dot(z, np.dot(y, x))
else:
raise Exception("image_dimension must be either 2 or 3.")
##
## Scale
##
scale = np.eye(params.image_dimension + 1)
scale_factor = np.random.uniform(1 - params.scale_range,
1 + params.scale_range)
for ax in params.img_spatial_indices:
scale[ax, ax] = scale_factor
##
## Flip
##
flip = np.eye(params.image_dimension + 1)
for f, ax in zip(params.flip, params.img_spatial_indices):
if f and (np.random.random() < 0.5):
flip[ax, ax] *= -1
##
## Compose
##
return np.dot(scale, np.dot(rotation, translation))
def zc_random(params, shape):
##
## Translation
##
translation = np.eye(params.image_dimension + 1)
for t, ax in zip(params.translation_range, params.img_spatial_indices):
random_t = np.random.uniform(t, t)
translation[ax, params.image_dimension] = random_t #* shape[ax]
if ax == 0:
print("translation_pixel:", random_t)
##
## Rotation
##
rotation = np.eye(params.image_dimension + 1)
if params.image_dimension == 2:
theta = (
np.pi / 180.0 *
np.random.uniform(-params.rotation_range, params.rotation_range))
rotation[:2, :2] = dv.rotation_matrix_from_angle(theta)
elif params.image_dimension == 3:
roll_degree = np.random.uniform(-params.rotation_range[0],
params.rotation_range[0])
print("roll_degree", roll_degree)
x = dv.rotation_matrix_from_angle(
np.pi / 180.0 * roll_degree,
matrix_type="roll",
)
pitch_degree = np.random.uniform(-params.rotation_range[1],
params.rotation_range[1])
print('pitch_degree', pitch_degree)
y = dv.rotation_matrix_from_angle(
np.pi / 180.0 * pitch_degree,
matrix_type="pitch",
)
yaw_degree = np.random.uniform(-params.rotation_range[2],
params.rotation_range[2])
print('yaw_degree', yaw_degree)
z = dv.rotation_matrix_from_angle(
np.pi / 180.0 * yaw_degree,
matrix_type="yaw",
)
rotation[:3, :3] = np.dot(z, np.dot(y, x))
else:
raise Exception("image_dimension must be either 2 or 3.")
##
## Scale
##
scale = np.eye(params.image_dimension + 1)
scale_factor = np.random.uniform(1 - params.scale_range,
1 + params.scale_range)
print('scale', scale_factor)
for ax in params.img_spatial_indices:
scale[ax, ax] = scale_factor
##
## Flip
##
flip = np.eye(params.image_dimension + 1)
for f, ax in zip(params.flip, params.img_spatial_indices):
if f and (np.random.random() < 0.5):
flip[ax, ax] *= -1
##
## Compose
##
return translation, rotation, scale, np.dot(scale,
np.dot(rotation, translation))
def test_rotation_matrix_from_angle():
print(dv.rotation_matrix_from_angle(np.pi / 2))
print(dv.rotation_matrix_from_angle(np.pi / 2, matrix_type="yaw"))
print(dv.rotation_matrix_from_angle(np.pi / 2, matrix_type="pitch"))
print(dv.rotation_matrix_from_angle(np.pi / 2, matrix_type="roll"))
def rotate_2d_vector(vec, ang):
return np.dot(dv.rotation_matrix_from_angle(ang), vec)