def default_camera(img_nr, img, H, scale, oshape, std=1.0, _coords=[]):
"""The default camera model simply blurs and downscales the image.
Parameters
----------
img_nr : int
The number of this image in the set. Useful for storing image-specific
parameters, such as coordinates.
img : ndarray
High-resolution image data.
H : (3,3) ndarray
Transformation matrix to apply to `img`.
oshape : tuple of ints
Output shape.
std : float
Standard deviation of the blur mask.
_coords : ndarray
Coordinates suitable for use in ``ndimage.map_coordinates``.
"""
try:
coords = _coords[img_nr]
except IndexError:
H = H.copy()
H[:2, :] *= float(scale)
H = np.linalg.inv(H)
coords = _homography_coords(img, H, oshape)
_coords.append(coords)
out = homography(img, H, _coords=coords)
out = out[:oshape[0], :oshape[1]]
return out
def default_camera(img_nr, img, H, scale, oshape, std=1.0, _coords=[]):
"""The default camera model simply blurs and downscales the image.
Parameters
----------
img_nr : int
The number of this image in the set. Useful for storing image-specific
parameters, such as coordinates.
img : ndarray
High-resolution image data.
H : (3,3) ndarray
Transformation matrix to apply to `img`.
oshape : tuple of ints
Output shape.
std : float
Standard deviation of the blur mask.
_coords : ndarray
Coordinates suitable for use in ``ndimage.map_coordinates``.
"""
try:
coords = _coords[img_nr]
except IndexError:
H = H.copy()
H[:2, :] *= float(scale)
H = np.linalg.inv(H)
coords = _homography_coords(img, H, oshape)
_coords.append(coords)
out = homography(img, H, _coords=coords)
out = out[:oshape[0], :oshape[1]]
return out
def cost(p, A, B):
M = _build_tf(p, translation_only=translation_only,
fixed_scale=fixed_scale)
try:
T = transform.homography(B, M, order=2)
except np.linalg.LinAlgError:
raise RuntimeError('Could not invert transformation matrix. '
'This may be because too many levels of '
'downscaling was requested (%d).' % levels)
H = joint_hist(A, T, win_size=win_size, std=std, fast=fast)
S = mutual_info(H)
return -S
def cost(p, A, B):
M = _build_tf(p, translation_only=translation_only,
fixed_scale=fixed_scale)
try:
T = transform.homography(B, M, order=2)
except np.linalg.LinAlgError:
raise RuntimeError('Could not invert transformation matrix. '
'This may be because too many levels of '
'downscaling was requested (%d).' % levels)
H = joint_hist(A, T, win_size=win_size, std=std, fast=fast)
S = mutual_info(H)
return -S
import numpy as np
from demo_data import chelsea
from supreme.register.parzen import joint_hist, mutual_info
from supreme.transform import homography
import matplotlib.pyplot as plt
h1 = chelsea(grey=True).astype(np.uint8)
plt.suptitle('Parzen-Window Joint PDF Estimator')
for n, t in enumerate([0, 0.01, 0.1]):
plt.subplot(1, 3, n + 1)
h2 = homography(h1, [[np.cos(t), -np.sin(t), 0],
[np.sin(t), np.cos(t), 0],
[0, 0, 1]])
if n == 0:
plt.ylabel('Grey levels in A')
plt.xlabel('Grey levels in B')
H = joint_hist(h1, h2, win_size=5, std=1)
S = mutual_info(H)
plt.title('Rotation $%.2f^\\circ$\nS=%.5f' % ((t / np.pi * 180), S))
plt.imshow(np.log(H + 10), interpolation='nearest')
plt.show()
import sys
if len(sys.argv) == 3:
from supreme.io import imread
A = imread(sys.argv[1], flatten=True).astype(np.uint8)
Ac = imread(sys.argv[1], flatten=False).astype(np.uint8)
B = imread(sys.argv[2], flatten=True).astype(np.uint8)
Bc = imread(sys.argv[2], flatten=False).astype(np.uint8)
else:
A = chelsea(grey=True).astype(np.uint8)
Ac = A
t = 20/180. * np.pi
x = 3
y = 6
B = homography(A, [[np.cos(t), -np.sin(t), x],
[np.sin(t), np.cos(t), y],
[0, 0, 1]])
Bc = B
M, S = register.dense_MI(A, B, levels=3, std=5, win_size=9)
print "Mutual information: ", S
if S < 1.5:
print "Warning: registration probably failed."
print "Transformation matrix:"
print np.array2string(M, separator=', ')
X = register.stack.with_transform([Ac, Bc], [np.eye(3), M], save_tiff=True)
if len(sys.argv) != 3:
import matplotlib.pyplot as plt
else:
parser.print_help()
sys.exit(0)
ic = load_vgg(vgg_dir)
# Perform crude photometric registration
ref = ic[0].copy()
images = []
scale_offset = []
for i in range(len(ic)):
scale = 1
offset = 0
if options.photo_adjust:
img_warp = homography(ic[i], ic[i].info['H'])
scale, offset = photometric_adjust(img_warp, ref)
scale_offset.append((scale, offset))
images.append(ic[i] * scale + offset)
print "Images adjusted by: %s" % str(['%.2f, %.2f' % (a,b)
for (a,b) in scale_offset])
images = [img for i,img in enumerate(images) if not i in options.ignore]
HH = [i.info['H'] for i in images]
oshape = np.floor(np.array(images[0].shape) * options.scale)
avg = initial_guess_avg(images, HH, options.scale, oshape)
#
import supreme.register as register
import sys
if len(sys.argv) == 3:
from supreme.io import imread
A = imread(sys.argv[1], flatten=True).astype(np.uint8)
Ac = imread(sys.argv[1], flatten=False).astype(np.uint8)
B = imread(sys.argv[2], flatten=True).astype(np.uint8)
Bc = imread(sys.argv[2], flatten=False).astype(np.uint8)
else:
A = chelsea(grey=True).astype(np.uint8)
Ac = A
t = 20 / 180. * np.pi
x = 3
y = 6
B = homography(
A, [[np.cos(t), -np.sin(t), x], [np.sin(t), np.cos(t), y], [0, 0, 1]])
Bc = B
M, S = register.dense_MI(A, B, levels=3, std=5, win_size=9)
print "Mutual information: ", S
if S < 1.5:
print "Warning: registration probably failed."
print "Transformation matrix:"
print np.array2string(M, separator=', ')
X = register.stack.with_transform([Ac, Bc], [np.eye(3), M], save_tiff=True)
if len(sys.argv) != 3:
import matplotlib.pyplot as plt
plt.imshow(X)
import numpy as np
from demo_data import chelsea
from supreme.register.parzen import joint_hist, mutual_info
from supreme.transform import homography
import matplotlib.pyplot as plt
h1 = chelsea(grey=True).astype(np.uint8)
plt.suptitle('Parzen-Window Joint PDF Estimator')
for n, t in enumerate([0, 0.01, 0.1]):
plt.subplot(1, 3, n + 1)
h2 = homography(
h1, [[np.cos(t), -np.sin(t), 0], [np.sin(t), np.cos(t), 0], [0, 0, 1]])
if n == 0:
plt.ylabel('Grey levels in A')
plt.xlabel('Grey levels in B')
H = joint_hist(h1, h2, win_size=5, std=1)
S = mutual_info(H)
plt.title('Rotation $%.2f^\\circ$\nS=%.5f' % ((t / np.pi * 180), S))
plt.imshow(np.log(H + 10), interpolation='nearest')
plt.show()
else:
parser.print_help()
sys.exit(0)
ic = load_vgg(vgg_dir)
# Perform crude photometric registration
ref = ic[0].copy()
images = []
scale_offset = []
for i in range(len(ic)):
scale = 1
offset = 0
if options.photo_adjust:
img_warp = homography(ic[i], ic[i].info['H'])
scale, offset = photometric_adjust(img_warp, ref)
scale_offset.append((scale, offset))
images.append(ic[i] * scale + offset)
print "Images adjusted by: %s" % str(
['%.2f, %.2f' % (a, b) for (a, b) in scale_offset])
images = [img for i, img in enumerate(images) if not i in options.ignore]
HH = [i.info['H'] for i in images]
oshape = np.floor(np.array(images[0].shape) * options.scale)
avg = initial_guess_avg(images, HH, options.scale, oshape)
#
