def align(img, landmarks, image_size=(112, 112)): ''' Takes oirignial image and detected landmarks as numpy format :param img: original image :param landmarks: detected landmarks, [[x1, y1], [x2, y2], [x3, y3], [x4, y4], [x5, y5]] :param image_size: output image_size :return warped_landmarks: :return warped_image: ''' assert isinstance(img, np.ndarray) assert isinstance(landmarks, np.ndarray) assert landmarks.shape == (5, 2) M = None src = np.array([ [30.2946, 51.6963], [65.5318, 51.5014], [48.0252, 71.7366], [33.5493, 92.3655], [62.7299, 92.2041]], dtype=np.float32) if image_size[1] == 112: src[:, 0] += 8.0 dst = landmarks.astype(np.float32) tform = SimilarityTransform() tform.estimate(dst, src) M = tform.params[0:2, :] warped_image = cv2.warpAffine(img, M, image_size, borderValue=0.0) warped_landmarks = cv2.perspectiveTransform(landmarks[np.newaxis, ...], tform.params[0:3, :])[0] return warped_landmarks, warped_image
def warp_images(image, transform, translation=None):
r, c = image.shape[:2]
# Note that transformations take coordinates in (x, y) format,
# not (row, column), in order to be consistent with most literature
corners = np.array([[0, 0], [0, r], [c, r], [c, 0]]).astype(np.float)
# Warp the image corners to their new positions
if translation is not None:
corners -= translation
warped_corners = transform(corners)
# Find the extents of both the reference image and the warped
# target image
all_corners = np.vstack((warped_corners))
corner_min = np.min(all_corners, axis=0)
corner_max = np.max(all_corners, axis=0)
output_shape = (corner_max - corner_min)
output_shape = np.ceil(output_shape[::-1])
translation_0 = np.array([0.0, 0.0])
if translation is not None:
translation_0 += translation
offset_0 = SimilarityTransform(translation=-translation_0)
offset = SimilarityTransform(translation=-corner_min)
image = warp(image, (offset_0 + transform + offset).inverse,
output_shape=output_shape,
cval=0)
im = Image.fromarray((image).astype('uint8'))
return im, offset.translation
def test_3d_similarity_estimation():
src_points = np.random.rand(1000, 3)
# Random transformation for testing
angles = np.random.random((3, )) * 2 * np.pi - np.pi
scale = np.random.randint(0, 20)
rotation_matrix = _euler_rotation_matrix(angles) * scale
translation_vector = np.random.random((3, ))
dst_points = []
for pt in src_points:
pt_r = pt.reshape(3, 1)
dst = np.matmul(rotation_matrix, pt_r) + \
translation_vector.reshape(3, 1)
dst = dst.reshape(3)
dst_points.append(dst)
dst_points = np.array(dst_points)
# estimating the transformation
tform = SimilarityTransform(dimensionality=3)
assert tform.estimate(src_points, dst_points)
estimated_rotation = tform.rotation
estimated_translation = tform.translation
estimated_scale = tform.scale
assert_almost_equal(estimated_translation, translation_vector)
assert_almost_equal(estimated_scale, scale)
assert_almost_equal(estimated_rotation, rotation_matrix)
def _rotate(image, angle, center, scale, cval=0):
"""
Rotate function taken mostly from scikit image. Main difference is that
this one allows dimensional scaling and records the final translation
to ensure no image content is lost. This is needed to rotate the seam
back into the original image.
"""
rows, cols = image.shape[0], image.shape[1]
tform1 = SimilarityTransform(translation=center)
tform2 = SimilarityTransform(rotation=angle)
tform3 = SimilarityTransform(translation=-center)
tform4 = AffineTransform(scale=(1 / scale, 1))
tform = tform4 + tform3 + tform2 + tform1
corners = np.array([[0, 0], [0, rows - 1], [cols - 1, rows - 1],
[cols - 1, 0]])
corners = tform.inverse(corners)
minc = corners[:, 0].min()
minr = corners[:, 1].min()
maxc = corners[:, 0].max()
maxr = corners[:, 1].max()
out_rows = maxr - minr + 1
out_cols = maxc - minc + 1
output_shape = np.around((out_rows, out_cols))
# fit output image in new shape
translation = (minc, minr)
tform5 = SimilarityTransform(translation=translation)
tform = tform5 + tform
tform.params[2] = (0, 0, 1)
return tform, warp(image,
tform,
output_shape=output_shape,
order=0,
cval=cval,
clip=False,
preserve_range=True)
def rotate_selection(self, event):
selection = self.get_selection()
sel_pos = np.array([self.canvas.coords(s) for s in selection])
sel_pos = sel_pos[..., :2] + (sel_pos[..., 2:] - sel_pos[..., :2]) / 2
sel_center = np.mean(sel_pos, axis=0)
if self.rotation_center is not None:
sel_center = self.rotation_center
pos_old = np.array([self._drag_data["x"], self._drag_data["y"]])
if np.all(pos_old == np.array([0, 0])):
pos_old = np.array([event.x, event.y])
direction_old = pos_old - sel_center
pos_new = np.array([event.x, event.y])
direction_new = pos_new - sel_center
X = np.array([direction_new[0], direction_old[0]])
Y = np.array([direction_new[1], direction_old[1]])
angle = np.arctan2(Y, X)
angle = angle[1] - angle[0]
rot_trans = SimilarityTransform(rotation=angle)
new_coords = rot_trans.inverse(sel_pos - sel_center) + sel_center
for s, old, new in zip(selection, sel_pos, new_coords):
delta_x = new[0] - old[0]
delta_y = new[1] - old[1]
self.canvas.move(s, delta_x, delta_y)
self._drag_data["x"] = event.x
self._drag_data["y"] = event.y
def check_ruler_points():
points = pd.read_csv('../output/ruler_points.csv')
for _, row in points.iterrows():
video_id = row.video_id
img = scipy.misc.imread(os.path.join(IMAGES_DIR, video_id, "0001.jpg"))
dst_w = 720
dst_h = 360
ruler_points = np.array([[row.ruler_x0, row.ruler_y0],
[row.ruler_x1, row.ruler_y1]])
img_points = np.array([[dst_w * 0.1, dst_h / 2],
[dst_w * 0.9, dst_h / 2]])
tform = SimilarityTransform()
tform.estimate(dst=ruler_points, src=img_points)
crop = skimage.transform.warp(img,
tform,
mode='edge',
order=3,
output_shape=(dst_h, dst_w))
print('ruler:\n', ruler_points)
print('img:\n', img_points)
print('ruler from img:\n', tform(img_points))
print('img from ruler:\n', tform.inverse(ruler_points))
print('scale', tform.scale)
plt.subplot(2, 1, 1)
plt.imshow(img)
plt.plot([row.ruler_x0, row.ruler_x1], [row.ruler_y0, row.ruler_y1])
plt.subplot(2, 1, 2)
plt.imshow(crop)
plt.show()
def transform_for_clip(self,
video_id,
dst_w=720,
dst_h=360,
points_random_shift=0):
"""Finds transform to crop around ruler.
# Arguments
video_id: Video ID.
dst_w: Width of cropped image.
dst_h: Height of cropped image.
points_random_shift: How many points to randomly shift image.
"""
img_points = np.array([[dst_w * 0.1, dst_h / 2],
[dst_w * 0.9, dst_h / 2]])
if video_id not in self.ruler_points:
return None
points = self.ruler_points[video_id]
ruler_points = np.array([[points.x1, points.y1],
[points.x2, points.y2]])
if points_random_shift > 0:
img_points += np.random.uniform(-points_random_shift,
points_random_shift, (2, 2))
tform = SimilarityTransform()
tform.estimate(dst=ruler_points, src=img_points)
return tform
def get_face(img, dst, target_size=(112, 112)):
"""
:param img: image
:param dst:
:param target_size:
:return:
"""
# MS1M
# [38.128662 51.516567]
# [74.21549 51.55989 ]
# [56.056564 72.434525]
# [40.48149 90.873665]
# [71.38436 90.78255 ]
# LFW
# [38.411846 52.59001 ]
# [73.68209 52.300644]
# [56.092415 72.949585]
# [40.763634 90.94648 ]
# [71.64599 90.62956 ]
src = np.array(
[
[38.2946, 51.6963],
[73.5318, 51.5014],
[56.0252, 71.7366],
[41.5493, 92.3655],
[70.7299, 92.2041],
],
dtype=np.float32,
)
tform = SimilarityTransform()
tform.estimate(dst, src)
tmatrix = tform.params[0:2, :]
return cv2.warpAffine(img, tmatrix, target_size, borderValue=0.0)
def test_similarity_init():
# init with implicit parameters
scale = 0.1
rotation = 1
translation = (1, 1)
tform = SimilarityTransform(scale=scale,
rotation=rotation,
translation=translation)
assert_array_almost_equal(tform.scale, scale)
assert_array_almost_equal(tform.rotation, rotation)
assert_array_almost_equal(tform.translation, translation)
# init with transformation matrix
tform2 = SimilarityTransform(tform._matrix)
assert_array_almost_equal(tform2.scale, scale)
assert_array_almost_equal(tform2.rotation, rotation)
assert_array_almost_equal(tform2.translation, translation)
# test special case for scale if rotation=0
scale = 0.1
rotation = 0
translation = (1, 1)
tform = SimilarityTransform(scale=scale,
rotation=rotation,
translation=translation)
assert_array_almost_equal(tform.scale, scale)
assert_array_almost_equal(tform.rotation, rotation)
assert_array_almost_equal(tform.translation, translation)
def align_face(self, img, src_img_size, landmark, origin):
dst_size = 112
new_dst_size = 224
dst = np.array([
[30.2946 + 8, 51.6963],
[65.5318 + 8, 51.5014],
[48.0252 + 8, 71.7366],
[33.5493 + 8, 92.3655],
[62.7299 + 8, 92.2041]], dtype=np.float32 ) * new_dst_size / dst_size
p = src_img_size / new_dst_size
dst = dst * p
src = landmark - np.array(origin)
# print("landmark2", src)
#dst = np.transpose(landmark).reshape(1,5,2)
#src = src.reshape(1,5,2)
# print(src)
# print(dst)
# transmat = cv2.estimateRigidTransform(dst.astype(np.float32),
# src.astype(np.float32), False)
# out = cv2.warpAffine(img, transmat, (dst_img_size, dst_img_size))
tform = SimilarityTransform()
tform.estimate(src, dst)
M = tform.params[0:2, :]
out = cv2.warpAffine(img, M, (src_img_size, src_img_size), borderValue=0.0)
return out
def doOpticalFlow(prevOutput, targetPoints, target_frame, prev_target_frame,
frame_no):
p0 = np.asarray(targetPoints).astype(np.float32)[:, :, None]
p0 = np.transpose(p0, (0, 2, 1))
old_gray = cv2.cvtColor(prev_target_frame, cv2.COLOR_BGR2GRAY)
frame_gray = cv2.cvtColor(target_frame, cv2.COLOR_BGR2GRAY)
p1, st, err = cv2.calcOpticalFlowPyrLK(old_gray, frame_gray, p0, None,
**lk_params)
# Select good points
good_new = p1[st == 1]
good_old = p0[st == 1]
newOutput = np.copy(prevOutput)
transform = SimilarityTransform()
if transform.estimate(good_old, good_new):
newOutput = transform_image(good_old, good_new, prevOutput,
target_frame, frame_no)
# make a convex fill mask inside this hull
# for each point in the new image, copy from the old image/prevOutput (use transformation/float points..)
# use the above mask to only keep the masked part of the above copy
# seamless cloning will not be required.
return newOutput, listOfListToTuples(good_new.tolist())
def im_affine_transform(img, scale, rotation, shear, translation_y, translation_x, return_tform=False):
# Assumed img in c01. Convert to 01c for skimage
img = img.transpose(1, 2, 0)
# Normalize so that the param acts more like im_rotate, im_translate etc
scale = 1 / scale
translation_x = - translation_x
translation_y = - translation_y
# shift to center first so that image is rotated around center
center_shift = np.array((img.shape[0], img.shape[1])) / 2. - 0.5
tform_center = SimilarityTransform(translation=-center_shift)
tform_uncenter = SimilarityTransform(translation=center_shift)
rotation = np.deg2rad(rotation)
tform = AffineTransform(scale=(scale, scale), rotation=rotation,
shear=shear,
translation=(translation_x, translation_y))
tform = tform_center + tform + tform_uncenter
warped_img = warp(img, tform)
# Convert back from 01c to c01
warped_img = warped_img.transpose(2, 0, 1)
warped_img = warped_img.astype(img.dtype)
if return_tform:
return warped_img, tform
else:
return warped_img
def merge_two_images(image0, offset0, image1, offset1):
cmin_0, cmax_0 = get_inplace_images_conners(image0, offset0)
cmin_1, cmax_1 = get_inplace_images_conners(image1, offset1)
out_min = np.min(np.vstack((cmin_0, cmin_1, cmax_0, cmax_1)), axis=0)
out_max = np.max(np.vstack((cmin_0, cmin_1, cmax_0, cmax_1)), axis=0)
print(out_min, out_max)
output_shape = (out_max - out_min)
output_shape = np.ceil(output_shape[::-1])
im0_translation = SimilarityTransform(translation=-offset_0)
im1_translation = SimilarityTransform(translation=-offset_1)
all_translation = SimilarityTransform(translation=-out_min)
image0 = warp(image0, (im0_translation + all_translation).inverse,
output_shape=output_shape,
cval=0)
image1 = warp(image1, (im1_translation + all_translation).inverse,
output_shape=output_shape,
cval=0)
im0 = Image.fromarray((image0).astype('uint8'))
im1 = Image.fromarray((image1).astype('uint8'))
im0.show()
im1.show()
def jitter(img):
''' Jitter the image as described in the paper referenced here:
http://yann.lecun.com/exdb/publis/pdf/sermanet-ijcnn-11.pdf'''
# knobs
max_rot = 15 * math.pi / 180.0
max_scale = 0.1
max_delta = 2
max_gamma = 0.7
# randomize
rot = random.uniform(-1, 1) * max_rot
scl = random.uniform(-1, 1) * max_scale + 1.0
xd = random.randint(-max_delta, max_delta)
yd = random.randint(-max_delta, max_delta)
gamma = random.uniform(-1, 1) * max_gamma + 1.0
# scale, roation, and translation
tform = SimilarityTransform(rotation=rot, scale=scl, translation=(xd, yd))
offx, offy = np.array(img.shape[:2]) / 2
recenter = SimilarityTransform(translation=(offx, offy))
recenter_inv = SimilarityTransform(translation=(-offx, -offy))
img = warp(img, (recenter_inv + (tform + recenter)).inverse, mode='edge')
# gamma
img = adjust_gamma(img, gamma)
# convert back to RGB [0-255] and ignore the silly precision warning
with warnings.catch_warnings():
warnings.simplefilter("ignore")
return img_as_ubyte(img)
def __call__(self, image):
"""
"""
downscaled = self.downscale(image) / 256
h, w = downscaled.shape[:2]
u, v = self.gcnet_shape
half_size = (min(h, w) - 1) // 2
scale = min(image.shape[:2]) / u
#print(half_size, u, v, h, w, scale)
src = np.array([
(w / 2 - half_size, h / 2 - half_size), # A
(w / 2 + half_size, h / 2 - half_size), # B
(w / 2 - half_size, h / 2 + half_size), # C
])
#print(src)
dst = np.array([
(0, 0), # A
(u, 0), # B
(0, v) # C
])
T = SimilarityTransform()
T.estimate(src, dst)
X = warp(downscaled, T.inverse, output_shape=self.gcnet_shape,
order=2).astype(np.float32)
T_pred, R_pred = self.gcnet.predict(X[None, ...])
return theta_from_soft_label(T_pred), scale * rho_from_soft_label(
R_pred, K_range=100)
def upgrade(self, mult):
# Creates a train data set via concatenating morphing initial images with slight rotation and shifting mult times.
# Returns a tuple of a train data set and a labels vector.
X_upg = np.array([
warp(
item,
SimilarityTransform(translation=(np.random.randint(-1, 1),
np.random.randint(-1, 1))))
for item in
[rotate(item, np.random.randint(-15, 15)) for item in self.X]
])
y_upg = self.y
for i in range(mult - 1):
X_upg = np.concatenate(
(X_upg,
np.array([
warp(
item,
SimilarityTransform(
translation=(np.random.randint(-1, 1),
np.random.randint(-1, 1))))
for item in [
rotate(item, np.random.randint(-15, 15))
for item in self.X
]
])))
y_upg = np.concatenate((y_upg, self.y))
print('iteration {0} is sucsessfully done'.format(i))
return X_upg, y_upg
def register_histo(histo_img, preprocessed_mr_images, ref='t1', n_points=7):
grayscale_histo = np.mean(histo_img, axis=2)
reference = nib.load(preprocessed_mr_images.file_paths[ref]).get_data()
myfig = plt.figure(4)
myfig.suptitle("Select corresponding landmarks (" + str(n_points) + ") for similarity transform !")
plt.subplot(1, 2, 1)
plt.imshow(grayscale_histo, cmap="gray")
x_histo = plt.ginput(n_points, timeout=0)
print("Selected points for histological cut:")
x_histo = np.array(x_histo)
print(x_histo)
plt.subplot(1, 2, 2)
plt.imshow(reference, cmap="gray")
x_mr = plt.ginput(n_points, timeout=0)
print("Selected points for " + ref + ":")
x_mr = np.array(x_mr)
print(x_mr)
sim_transform = SimilarityTransform()
sim_transform.estimate(x_mr, x_histo)
print(Fore.GREEN + "Parameter estimation - transformation matrix: " + Fore.RESET)
print(sim_transform.params)
warped = warp(grayscale_histo,
sim_transform,
output_shape=reference.shape)
# grayscale_histo = grayscale_histo.astype('float32')
return sim_transform.params, warped
def __image_transform__(self,theImage,scaleFactor,rotationAngle,txFactor,tyFactor):
centerY,centerX=np.array(theImage.shape[:2])/2.
theRotation=SimilarityTransform(rotation=np.deg2rad(rotationAngle))
theZoom=SimilarityTransform(scale=scaleFactor)
theShift=SimilarityTransform(translation=[-centerX,-centerY])
theShiftInv=SimilarityTransform(translation=[centerX,centerY])
theTranslation=SimilarityTransform(translation=[txFactor*2*centerX,tyFactor*2*centerY])
return warp(theImage, (theShift+(theRotation+theShiftInv))+(theShift+(theZoom+theShiftInv))+theTranslation, mode='reflect')
def test_union():
tform1 = SimilarityTransform(scale=0.1, rotation=0.3)
tform2 = SimilarityTransform(scale=0.1, rotation=0.9)
tform3 = SimilarityTransform(scale=0.1 ** 2, rotation=0.3 + 0.9)
tform = tform1 + tform2
assert_array_almost_equal(tform._matrix, tform3._matrix)
def create_img_mask(r, c, n, transforms, priorities):
pro_idx = priorities[0] # [0, 1, 2, 3]
print(r, c, n)
stitch_masks = np.array([pow(2, i) * np.ones((r, c)) for i in range(2)])
return_masks = []
corners = np.array([[0, 0], [0, r], [c, r], [c, 0]]).astype(np.float)
for index in range(n):
# create mask for index-th image
if index == pro_idx:
return_masks.append(None)
continue
else:
al_corners = corners
warped_corners = transforms[pro_idx](corners)
al_corners = np.vstack((al_corners, warped_corners))
warped_corners = transforms[index](corners)
al_corners = np.vstack((al_corners, warped_corners))
corner_min = np.min(al_corners, axis=0)
corner_max = np.max(al_corners, axis=0)
output_shape = (corner_max - corner_min)
output_shape = np.ceil(output_shape[::-1])
offset = SimilarityTransform(translation=-corner_min)
offset_inv = SimilarityTransform(translation=corner_min)
total_masks = []
total_masks.append(
warp(stitch_masks[pro_idx, :, :],
(transforms[pro_idx] + offset).inverse,
output_shape=output_shape,
cval=0))
total_masks.append(
warp(stitch_masks[1, :, :], (transforms[1] + offset).inverse,
output_shape=output_shape,
cval=0))
total_masks = np.sum(np.array(total_masks), axis=0)
# return val
transform_inv = ProjectiveTransform(transforms[index]._inv_matrix)
return_masks.append(
warp(total_masks, (offset_inv + transform_inv).inverse,
output_shape=[r, c],
cval=0))
return_masks[index][(return_masks[index] % 1.0 != 0)] = pow(
2, 1) # pow(2,i)
ret_masks = return_masks[index]
# now the image that has to be bitwise-and. so the background image has to be 255
# the mask[i]. the overlap_label will be 2^len - 1 = 3, overlap label
overlap_value = pow(2, 2) - 1 # 3
# print ((ret_masks==3.0).sum())
ret_masks[(ret_masks != overlap_value)] = 255 # white
ret_masks[(ret_masks == overlap_value)] = 0 # reverse mask
ret_masks = ret_masks.astype('uint8')
# print((ret_masks[index] == 255).sum())
return return_masks
def get_pairs(imgs, unique_pairs, offsets, subsample_factor, overlap_pixels,
n_kp):
"""Create inlier keypoint pairs."""
orb = ORB(n_keypoints=n_kp, fast_threshold=0.05)
k = gaussian(offsets * 2 + 1, 1, sym=True)
tf = SimilarityTransform # tf = RigidTransform
tf0 = SimilarityTransform()
# FIXME: is there no rigid model in scikit-image???
# this is needed for input to RANSAC
pairs = []
init_tfs = np.empty([n_slcs, n_tiles, 3])
for p in unique_pairs:
pair_tstart = time()
full_im1, full_im2 = subsample_images(p, imgs, subsample_factor)
part_im1, part_im2 = select_imregions(p[2], full_im1, full_im2,
overlap_pixels)
keyp_im1, desc_im1 = get_keypoints(orb, part_im1)
keyp_im2, desc_im2 = get_keypoints(orb, part_im2)
keyp_im1, keyp_im2 = reset_imregions(p[2], keyp_im1, keyp_im2,
overlap_pixels, full_im1.shape)
matches = match_descriptors(desc_im1, desc_im2, cross_check=True)
dst = keyp_im1[matches[:, 0]][:, ::-1]
src = keyp_im2[matches[:, 1]][:, ::-1]
model, inliers = ransac((src, dst),
tf,
min_samples=4,
residual_threshold=2,
max_trials=300)
w = k[offsets - (p[1][0] - p[0][0])]
pairs.append((p, src[inliers], dst[inliers], model, w))
if (p[0][1] == 0) & (p[0][0] == p[1][0]
): # referenced to tile 0 within the same slice
tf1 = tf0.__add__(model)
itf = [
math.acos(min(tf1.params[0, 0],
1)), # FIXME!!! with RigidTransform
tf1.params[0, 2],
tf1.params[1, 2]
]
init_tfs[p[1][0], p[1][1], :] = np.array(itf)
if (p[0][1] == p[1][1] == 0) & (
p[1][0] - p[0][0] == 1): # if [slcX,tile0] to [slcX-1,tile0]
tf0 = tf0.__add__(model)
plot_pair_ransac(p, full_im1, full_im2, keyp_im1, keyp_im2, matches,
inliers)
print('Pair done in: %.2f s' % (time() - pair_tstart, ))
return pairs, init_tfs
def test_zero_image_size():
with testing.raises(ValueError):
warp(np.zeros(0), SimilarityTransform())
with testing.raises(ValueError):
warp(np.zeros((0, 10)), SimilarityTransform())
with testing.raises(ValueError):
warp(np.zeros((10, 0)), SimilarityTransform())
with testing.raises(ValueError):
warp(np.zeros((10, 10, 0)), SimilarityTransform())
def build_center_uncenter_transforms(image_shape):
"""
These are used to ensure that zooming and rotation happens around the center of the image.
Use these transforms to center and uncenter the image around such a transform.
Copied this function from Kaggle NDSB winners: https://github.com/benanne/kaggle-ndsb
"""
center_shift = np.array([image_shape[1], image_shape[0]]) / 2.0 - 0.5 # need to swap rows and cols here apparently! confusing!
tform_uncenter = SimilarityTransform(translation=-center_shift)
tform_center = SimilarityTransform(translation=center_shift)
return tform_center, tform_uncenter
def transform_image(image, rotation=0):
transform = AffineTransform(rotation=np.deg2rad(rotation))
shift_y, shift_x = (np.array(image.shape) - 1) / 2.
shift_fwd = SimilarityTransform(translation=[-shift_x, -shift_y])
shift_back = SimilarityTransform(translation=[shift_x, shift_y])
transformed = warp(image, (shift_fwd + (transform + shift_back)).inverse,
order=1,
preserve_range=True,
mode='constant')
return transformed
def rotate(image, angle, cval):
center = np.array((image.shape[0], image.shape[1])) / 2. - 0.5
tform1 = SimilarityTransform(translation=center)
tform2 = SimilarityTransform(rotation=angle * np.pi / 180)
tform3 = SimilarityTransform(translation=-center)
tform = tform3 + tform2 + tform1
result = []
for i in range(image.shape[2]):
result.append(warp(image[:, :, i], tform, cval=cval))
return np.transpose(result, (1, 2, 0))
def scale_image(img, scaling_factor):
"""Scale the image foreground
This function scales the image foreground by a factor.
The background of the image is assumed to be black (0 grayscale).
.. versionadded:: 0.7
Parameters
----------
img : ndarray
A 2D NumPy array representing a (height, width) grayscale image, or a
3D NumPy array representing a (height, width, channels) RGB image
scaling_factor : float
Factory by which to scale the image
Returns
-------
img : ndarray
A copy of the scaled image
"""
if img.ndim < 2 or img.ndim > 3:
raise ValueError("Only 2D and 3D images are allowed, not "
"%dD." % img.ndim)
if img.ndim == 3 and img.shape[-1] > 3:
raise ValueError("Only RGB and grayscale images are allowed, not "
"%d-channel images." % img.shape[-1])
if scaling_factor <= 0:
raise ValueError("Scaling factor must be greater than zero")
# Calculate center of mass:
m = moments(img, order=1)
# No area found:
if isclose(m[0, 0], 0):
return img
# Shift the phosphene to (0, 0):
center_mass = np.array([m[0, 1] / m[0, 0], m[1, 0] / m[0, 0]])
tf_shift = SimilarityTransform(translation=-center_mass)
# Scale the phosphene:
tf_scale = SimilarityTransform(scale=scaling_factor)
# Shift the phosphene back to where it was:
tf_shift_inv = SimilarityTransform(translation=center_mass)
# Combine all three transforms:
tf = tf_shift + tf_scale + tf_shift_inv
img_warped = warp(img, tf.inverse)
# Warp automatically converts to double, so we need to convert the image
# back to its original format:
if img.dtype == bool:
return img_as_bool(img_warped)
if img.dtype == np.uint8:
return img_as_ubyte(img_warped)
if img.dtype == np.float32:
return img_as_float32(img_warped)
return img_warped
def distort(img):
shift_y, shift_x = np.array(img.shape[:2]) / 2.
shift = SimilarityTransform(translation=[-shift_x, -shift_y])
tf = SimilarityTransform(
rotation=np.deg2rad(random.uniform(-5.0, 5.0)),
scale=random.uniform(0.9, 1.1),
translation=(random.uniform(-0.1, 0.1)*img.shape[0], random.uniform(-0.1, 0.1)*img.shape[1])
)
shift_inv = SimilarityTransform(translation=[shift_x, shift_y])
return warp(img, (shift + (tf + shift_inv)).inverse, mode='edge')
def align_face(img, landmark, destination_size=(96, 112)):
# TODO :: 원하는 사이즈의 이미지로 align
"""
얼굴 랜드마크 좌표 (눈, 코, 입)을 중심으로 이미지를 정렬함.
:param img: np array image (H, W, C)
:param landmark: np array coordinates 5 x 2. 눈1, 눈2, 코, 입 양 끝.
if len(landmark) == 68:
It will be converted to 5x2
68개 랜드마크 기준으로
[36:42 평균,
42:48 평균,
30,
48,
54]
:return:
"""
if len(landmark) == 68:
landmark = np.array(landmark)
landmark = np.array([
landmark[36:42, :2].mean(axis=0), landmark[42:48, :2].mean(axis=0),
landmark[30, :2], landmark[48, :2], landmark[54, :2]
])
ref_size = np.array([96, 112])
# dst = np.array([
# [30.2946, 51.6963],
# [65.5318, 51.5014],
# [48.0252, 71.7366],
# [33.5493, 92.3655],
# [62.7299, 92.2041]], dtype=np.float32)
dst = np.array([[30.2946, 51.6963], [65.5318, 51.5014], [48.0252, 71.7366],
[33.5493, 92.3655], [62.7299, 92.2041]],
dtype=np.float32)
dst_size = np.array(destination_size)
p = dst_size / ref_size
dst = dst * p
src = landmark
tform = SimilarityTransform()
tform.estimate(src, dst)
M = tform.params[0:2, :]
out = cv2.warpAffine(img, M, destination_size, borderValue=0.0)
return out
def apply_random_transformation(background_size, segmented_box, config_params):
"""apply a random transformation to 2D coordinates nomalized to image size"""
# translate object coordinates to the object center's frame, i.e. whitens
whitened_coords_norm = segmented_box.segmented_coords_norm - (
segmented_box.x_center_norm, segmented_box.y_center_norm)
# then generate a random rotation around the z-axis (perpendicular to the image plane), and limit the object scale
# to maximum (default) 50% of the background image, i.e. the normalized largest dimension of the object must be at
# most 0.5. To put it simply, scale the objects down if they're too big.
# TODO(minhnh) add shear
max_scale = config_params.max_scale
if segmented_box.max_dimension_norm > config_params.max_obj_size_in_bg:
max_scale = config_params.max_obj_size_in_bg / segmented_box.max_dimension_norm
random_rot_angle = np.random.uniform(0, np.pi)
rand_scale = np.random.uniform(config_params.min_scale,
config_params.max_scale)
# generate a random translation within the image boundaries for whitened, normalized coordinates, taking into
# account the maximum allowed object dimension. After this translation, the normalized coordinates should
# stay within [margin, 1-margin] for each dimension
scaled_max_dimension = segmented_box.max_dimension_norm * max_scale
low_norm_bound, high_norm_bound = ((scaled_max_dimension / 2) +
config_params.margin,
1 - config_params.margin -
(scaled_max_dimension / 2))
random_translation_x = np.random.uniform(
low_norm_bound, high_norm_bound) * background_size[1]
random_translation_y = np.random.uniform(
low_norm_bound, high_norm_bound) * background_size[0]
# create the transformation matrix for the generated rotation, translation and scale
if np.random.uniform() < config_params.prob_rand_rotation:
tf_matrix = SimilarityTransform(
rotation=random_rot_angle,
scale=rand_scale * min(background_size),
translation=(random_translation_x, random_translation_y)).params
else:
tf_matrix = SimilarityTransform(
scale=rand_scale * min(background_size),
translation=(random_translation_x, random_translation_y)).params
# apply transformation
transformed_coords = matrix_transform(whitened_coords_norm, tf_matrix)
# we clip the object coordinates so that they are within the image boundaries
transformed_coords[np.where(transformed_coords[:, 0] < 0), 0] = 0
transformed_coords[np.where(transformed_coords[:, 0] > background_size[1]),
0] = background_size[1] - 1
transformed_coords[np.where(transformed_coords[:, 1] < 0), 1] = 0
transformed_coords[np.where(transformed_coords[:, 1] > background_size[0]),
1] = background_size[0] - 1
return transformed_coords
def augmentation_image_ms(image,
rotation_range=0,
shear_range=0,
scale_range=1,
transform_range=0,
horizontal_flip=False,
vertical_flip=False,
warp_mode='edge'):
from skimage.transform import AffineTransform, SimilarityTransform, warp
from numpy import deg2rad, flipud, fliplr
from numpy.random import uniform, random_integers
from random import choice
image_shape = image.shape
# Generate image transformation parameters
rotation_angle = uniform(low=-abs(rotation_range),
high=abs(rotation_range))
shear_angle = uniform(low=-abs(shear_range), high=abs(shear_range))
scale_value = uniform(low=abs(1 / scale_range), high=abs(scale_range))
translation_values = (random_integers(-abs(transform_range),
abs(transform_range)),
random_integers(-abs(transform_range),
abs(transform_range)))
# Horizontal and vertical flips
if horizontal_flip:
# randomly flip image up/down
if choice([True, False]):
image = flipud(image)
if vertical_flip:
# randomly flip image left/right
if choice([True, False]):
image = fliplr(image)
# Generate transformation object
transform_toorigin = SimilarityTransform(scale=(1, 1),
rotation=0,
translation=(-image_shape[0],
-image_shape[1]))
transform_revert = SimilarityTransform(scale=(1, 1),
rotation=0,
translation=(image_shape[0],
image_shape[1]))
transform = AffineTransform(scale=(scale_value, scale_value),
rotation=deg2rad(rotation_angle),
shear=deg2rad(shear_angle),
translation=translation_values)
# Apply transform
image = warp(image, ((transform_toorigin) + transform) + transform_revert,
mode=warp_mode,
preserve_range=True)
return image
def test_similarity_estimation():
# exact solution
tform = estimate_transform('similarity', SRC[:2, :], DST[:2, :])
assert_array_almost_equal(tform(SRC[:2, :]), DST[:2, :])
assert_equal(tform._matrix[0, 0], tform._matrix[1, 1])
assert_equal(tform._matrix[0, 1], - tform._matrix[1, 0])
# over-determined
tform2 = estimate_transform('similarity', SRC, DST)
assert_array_almost_equal(tform2.inverse(tform2(SRC)), SRC)
assert_equal(tform2._matrix[0, 0], tform2._matrix[1, 1])
assert_equal(tform2._matrix[0, 1], - tform2._matrix[1, 0])
# via estimate method
tform3 = SimilarityTransform()
tform3.estimate(SRC, DST)
assert_array_almost_equal(tform3._matrix, tform2._matrix)
def test_degenerate():
src = dst = np.zeros((10, 2))
tform = SimilarityTransform()
tform.estimate(src, dst)
assert np.all(np.isnan(tform.params))
tform = AffineTransform()
tform.estimate(src, dst)
assert np.all(np.isnan(tform.params))
tform = ProjectiveTransform()
tform.estimate(src, dst)
assert np.all(np.isnan(tform.params))
def test_union_differing_types():
tform1 = SimilarityTransform()
tform2 = PolynomialTransform()
with testing.raises(TypeError):
tform1.__add__(tform2)