-
Notifications
You must be signed in to change notification settings - Fork 2
/
mask_creator.py
271 lines (225 loc) · 13.3 KB
/
mask_creator.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
import cv2
import numpy as np
import os
import time
import shutil
class AugmentedDataset():
def __init__(self, root):
self.root = root
# self.imgs_backgrounds = list(sorted(os.listdir(os.path.join(root, 'backgrounds'))))
self.imgs_backgrounds = list(
sorted([f for f in os.listdir(os.path.join(root, 'backgrounds')) if not f.startswith('.')]))
self.imgs_templates = list(
sorted([f for f in os.listdir(os.path.join(root, 'templates')) if not f.startswith('.')]))
self.template_masks = self.create_template_masks(self.imgs_templates)
# number of max copies of the same template in one augmented image
self.max_templates = 3
# maximum relation between background to template
# the larger this value, the smaller is the maximum template size relative to its background
self.max_temp_back_rel = 20
# min scale of template when scaling the image down
# the larger this value, the larger is the minimum template size relative to its background
self.min_augm_scale = 0.7
# max rotation angle in the augmentation
self.max_augm_rot = 20.0
self.max_augm_rot_tem = 40.0
# max and min values (in percent of original) for illumination augmentation
self.min_illum = 0.2
self.max_illum = 2.5
# maximum perspective in given direction (must be 0<= x < 0.5)
# the smaller, the lesst perspective (0.0 means no added perspective)
self.max_perspective = 0.2
# maximum blur for image augmentation
self.max_blur = 3.0
def create_template_masks(self, imgs_templates):
template_masks = []
for template_name in imgs_templates:
template = cv2.imread(os.path.join(self.root, 'templates', template_name), cv2.IMREAD_UNCHANGED)
alpha_channel = template[:, :, 3]
_, template_mask = cv2.threshold(alpha_channel, 254, 1, cv2.THRESH_BINARY)
template_mask = template_mask.reshape((alpha_channel.shape[0], alpha_channel.shape[1]))
template_masks.append(template_mask)
return template_masks
def get_random_background(self):
idx = np.random.randint(0, len(self.imgs_backgrounds))
return self.imgs_backgrounds[idx]
def augment_templates(self, template, template_mask, amount):
augmented_templates = []
augmented_template_masks = []
for i in range(amount):
scale = np.random.uniform(self.min_augm_scale, 1.0)
template = cv2.resize(template, dsize=(0, 0), fx=scale, fy=scale)
template_mask = cv2.resize(template_mask, dsize=(0, 0), fx=scale, fy=scale, interpolation=cv2.INTER_AREA)
angle = np.random.uniform(-self.max_augm_rot_tem, self.max_augm_rot_tem)
template = self.rotate_image_keep_size(template, angle)
template_mask = self.rotate_image_keep_size(template_mask, angle)
horizontal_perspective = np.random.uniform(-self.max_perspective, self.max_perspective)
vertical_perspective = np.random.uniform(-self.max_perspective, self.max_perspective)
template = self.perspective_transformation(template, horizontal_perspective, vertical_perspective)
template_mask = self.perspective_transformation(template_mask, horizontal_perspective, vertical_perspective)
augmented_templates.append(template)
augmented_template_masks.append(template_mask)
return augmented_templates, augmented_template_masks
def stitch_templates_to_background(self, background, templates, template_masks, temp_count):
back_height = np.shape(background)[0]
back_width = np.shape(background)[1]
mask_array = [np.zeros_like(background[:, :, 0]) for _ in range(len(templates))]
for k in range(len(templates)):
# if len(templates) > 0:
rand_idx_array = np.random.permutation(len(templates[k]))
# print(np.shape(templates))
col = 1
for idx in rand_idx_array:
template = templates[k][idx]
template_mask = template_masks[k][idx]
temp_height = np.shape(template)[0]
temp_width = np.shape(template)[1]
y_coord = np.random.randint(0, back_height - temp_height)
x_coord = np.random.randint(0, back_width - temp_width)
y1, y2 = y_coord, y_coord + temp_height
x1, x2 = x_coord, x_coord + temp_width
for y in range(temp_height):
for x in range(temp_width):
if template_mask[y, x] > 0:
mask_array[k][y1 + y, x1 + x] = template_mask[y, x] * col
# mask_array[y1:y2, x1:x2] = template_mask * count_temp_cumsum[i]
col += 1
background_mask = 1.0 - template_mask
for c in range(0, 3):
background[y1:y2, x1:x2, c] = (
template_mask * template[:, :, c] + background_mask * background[y1:y2, x1:x2, c])
return background, mask_array
def rotate_image_keep_size(self, img, degreesCCW, scaleFactor=1):
(oldY, oldX) = img.shape[:2] # note: numpy uses (y,x) convention but most OpenCV functions use (x,y)
M = cv2.getRotationMatrix2D(center=(oldX / 2, oldY / 2), angle=degreesCCW,
scale=scaleFactor) # rotate about center of image.
# choose a new image size.
newX, newY = oldX * scaleFactor, oldY * scaleFactor
# include this if you want to prevent corners being cut off
r = np.deg2rad(degreesCCW)
newX, newY = (abs(np.sin(r) * newY) + abs(np.cos(r) * newX), abs(np.sin(r) * newX) + abs(np.cos(r) * newY))
# the warpAffine function call, below, basically works like this:
# 1. apply the M transformation on each pixel of the original image
# 2. save everything that falls within the upper-left "dsize" portion of the resulting image.
# So I will find the translation that moves the result to the center of that region.
(tx, ty) = ((newX - oldX) / 2, (newY - oldY) / 2)
M[0, 2] += tx # third column of matrix holds translation, which takes effect after rotation.
M[1, 2] += ty
rotatedImg = cv2.warpAffine(img, M, dsize=(int(newX), int(newY)), flags=cv2.INTER_LANCZOS4)
return rotatedImg
def rotate_image(self, image, angle, inter=cv2.INTER_LANCZOS4):
image_center = tuple(np.array(image.shape[1::-1]) / 2)
rot_mat = cv2.getRotationMatrix2D(image_center, angle, 1.0)
result = cv2.warpAffine(image, rot_mat, image.shape[1::-1], flags=inter)
return result
def change_illumination(self, image, gamma):
# changes illumination based on gamma correction: https://en.wikipedia.org/wiki/Gamma_correction
invGamma = 1.0 / gamma
table = np.array([((i / 255.0) ** invGamma) * 255
for i in np.arange(0, 256)]).astype("uint8")
return cv2.LUT(image, table)
def perspective_transformation(self, template, horizontal_perspective, vertical_perspective):
rows, cols = template.shape[:2]
src_points = np.float32([[0, 0], [cols - 1, 0], [0, rows - 1], [cols - 1, rows - 1]])
if (np.sign(horizontal_perspective) < 0.0) & (np.sign(vertical_perspective) < 0.0):
dst_points = np.float32(
[[int(np.abs(horizontal_perspective) / 2 * cols), int(np.abs(vertical_perspective) / 2 * rows)],
[int((1 - np.abs(horizontal_perspective) / 2) * cols), 0],
[0, int((1 - np.abs(vertical_perspective) / 2) * rows)],
[cols - 1, rows - 1]])
elif (np.sign(horizontal_perspective) >= 0.0) & (np.sign(vertical_perspective) < 0.0):
dst_points = np.float32(
[[0, int(np.abs(vertical_perspective) / 2 * rows)],
[cols - 1, 0],
[int(np.abs(horizontal_perspective) / 2 * cols), int((1 - np.abs(vertical_perspective) / 2) * rows)],
[int((1 - np.abs(horizontal_perspective) / 2) * cols), rows - 1]])
elif (np.sign(horizontal_perspective) < 0.0) & (np.sign(vertical_perspective) >= 0.0):
dst_points = np.float32(
[[int(np.abs(horizontal_perspective) / 2 * cols), 0],
[int((1 - np.abs(horizontal_perspective) / 2) * cols), int(np.abs(vertical_perspective) / 2 * rows)],
[0, rows - 1],
[cols - 1, int((1 - np.abs(vertical_perspective) / 2) * rows)]])
else:
dst_points = np.float32(
[[0, 0],
[cols - 1, int(np.abs(vertical_perspective) / 2 * rows)],
[int(np.abs(horizontal_perspective) / 2 * cols), rows - 1],
[int((1 - np.abs(horizontal_perspective) / 2) * cols),
int((1 - np.abs(vertical_perspective) / 2) * rows)]])
projective_matrix = cv2.getPerspectiveTransform(src_points, dst_points)
aug_template = cv2.warpPerspective(template, projective_matrix, (cols, rows), flags=cv2.INTER_LANCZOS4)
return aug_template
def blurr_image(self, image, sigma):
image = cv2.GaussianBlur(image, (0, 0), sigmaX=sigma)
return image
def get_train_data(self, amount):
aug_imgs = []
aug_masks = []
# data augmentation for background, template, and template mask
for i in range(amount):
print("Image: {}".format(i))
background_name = self.get_random_background()
background = cv2.imread(os.path.join(self.root, 'backgrounds', background_name), cv2.IMREAD_UNCHANGED)
background_size = background.shape[0] * background.shape[1]
stitch_templates = []
stitch_template_masks = []
temp_count = []
for j in range(len(self.imgs_templates)):
template = cv2.imread(os.path.join(self.root, 'templates', self.imgs_templates[j]),
cv2.IMREAD_UNCHANGED)
template_size = template.shape[0] * template.shape[1]
temp_back_rel = template_size / background_size
temp_normalization = 1 / np.sqrt(temp_back_rel * self.max_temp_back_rel)
template = cv2.resize(template, dsize=(0, 0), fx=temp_normalization, fy=temp_normalization,
interpolation=cv2.INTER_AREA)
template_mask = cv2.resize(self.template_masks[j], dsize=(0, 0), fx=temp_normalization,
fy=temp_normalization, interpolation=cv2.INTER_AREA)
rand = np.random.randint(1, self.max_templates + 1)
if rand > 0:
augmented_templates, augmented_template_masks = self.augment_templates(template, template_mask,
rand)
stitch_templates.append(augmented_templates)
stitch_template_masks.append(augmented_template_masks)
temp_count.append(rand)
# print(rand)
background_angle = np.random.uniform(-self.max_augm_rot, self.max_augm_rot)
# background = self.rotate_image(aug_mask[k], background_angle)
aug_img, aug_mask = self.stitch_templates_to_background(background, stitch_templates, stitch_template_masks,
temp_count)
# rotate the stitched images and masks for rotation augmentation
for k in range(len(aug_mask)):
aug_mask[k] = self.rotate_image(aug_mask[k], background_angle, cv2.INTER_LINEAR)
aug_img = self.rotate_image(aug_img, background_angle)
# change the illumination of the stitches images for illumination augmentation
gamma = np.random.uniform(self.min_illum, self.max_illum)
aug_img = self.change_illumination(aug_img, gamma=gamma)
# introduce Gaussian blur to the stitches images for blur augmentation
sigma = np.random.uniform(self.max_blur)
aug_img = self.blurr_image(aug_img, sigma)
cv2.imwrite("dataset/images/{}.png".format(i), aug_img)
os.mkdir("dataset/masks/{}".format(i))
# os.mkdir("dataset/visible_masks/{}".format(i))
# uncomment to show images created
# aug_imgs.append(aug_img)
# aug_masks.append(aug_mask)
# cv2.imshow('image', aug_img)
# cv2.waitKey()
for k, aug_m in enumerate(aug_mask):
cv2.imwrite("dataset/masks/{}/{}.png".format(i, k), aug_m)
# uncomment to show images and masks created
# mask = cv2.threshold(aug_m, 0, 255, cv2.THRESH_BINARY)
# cv2.imshow('mask', mask[1])
# cv2.waitKey()
# cv2.imwrite("dataset/visible_masks/{}/{}.png".format(i, k), mask[1])
# return aug_imgs, aug_masks
start = time.time()
np.random.seed(13245)
shutil.rmtree("dataset")
os.mkdir("dataset")
os.mkdir("dataset/images")
os.mkdir("dataset/masks")
# os.mkdir("dataset/visible_masks")
dataset = AugmentedDataset(os.path.join(os.path.dirname(os.path.abspath(__file__)), 'images'))
dataset.get_train_data(5)
end = time.time()
print(end - start)