/
augmentation.py
175 lines (124 loc) · 4.58 KB
/
augmentation.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
from PIL import Image
import torchvision.transforms as transformer
import random
import numpy as np
class Compose(object):
def __init__(self, transforms):
self.transforms = transforms
def __call__(self, image, targets):
for trans in self.transforms:
image, targets = trans(image, targets)
return image, targets
class Normalize(object):
def __call__(self, image, targets):
image = transformer.Normalize((0.5,0.5,0.5),(0.5,0.5,0.5))(image)
for i, target in enumerate(targets):
targets[i] = transformer.Normalize((0.5,),(0.5,))(target)
return image, targets
class ToTensor(object):
def __call__(self, image, targets):
image = transformer.ToTensor()(image)
for i, target in enumerate(targets):
targets[i] = transformer.ToTensor()(target)
return image, targets
class HFlip(object):
def __init__(self, p):
self.p = p
def __call__(self, image, targets):
if random.random() < self.p:
image = image.transpose(Image.FLIP_LEFT_RIGHT)
for i, target in enumerate(targets):
targets[i] = target.transpose(Image.FLIP_LEFT_RIGHT)
return image, targets
class ColorJitter(object):
def __call__(self, image, targets):
jitter_amount = 0.2
image = transformer.ColorJitter(jitter_amount, jitter_amount, jitter_amount, jitter_amount)(image)
for i, target in enumerate(targets):
targets[i] = transformer.ColorJitter(jitter_amount, jitter_amount, jitter_amount, jitter_amount)(target)
return image, targets
class Resize(object):
def __init__(self, output_size):
assert isinstance(output_size, tuple)
self.output_size = output_size
def __call__(self, image, targets):
image = image.resize(self.output_size, Image.BICUBIC)
for i, target in enumerate(targets):
targets[i] = target.resize(self.output_size, Image.BICUBIC)
return image, targets
class Crop(object):
def __init__(self, fine_size):
self.fine_size = fine_size
def __call__(self, image, targets):
#image = transformer.ToTensor()(image)
h, w = image.shape[:2]
w_offset = random.randint(0, max(0, h - self.fine_size - 1))
h_offset = random.randint(0, max(0, w - self.fine_size - 1))
image = image[:, h_offset:h_offset + self.fine_size,
w_offset:w_offset + self.fine_size]
for i, target in enumerate(targets):
targets[i] = target[:, h_offset:h_offset + self.fine_size,
w_offset:w_offset + self.fine_size]
return image, targets
class Rotation_and_Crop(object):
def __init__(self, p):
self.p = p
def __call__(self, image, targets):
if random.random() < self.p:
rot_deg = 5 * random.randint(-3, 3)
image = rotate_and_crop(image, rot_deg, True)
for i, target in enumerate(targets):
targets[i] = rotate_and_crop(target, rot_deg, True)
return image, targets
def perp(a):
# https://stackoverflow.com/questions/3252194/numpy-and-line-intersections
b = np.empty_like(a)
b[0] = -a[1]
b[1] = a[0]
return b
def seg_intersect(a1, a2, b1, b2):
# https://stackoverflow.com/questions/3252194/numpy-and-line-intersections
da = a2 - a1
db = b2 - b1
dp = a1 - b1
dap = perp(da)
denom = np.dot(dap, db)
num = np.dot(dap, dp)
return (num / denom.astype(float)) * db + b1
def rotate_and_crop(img, deg, same_size=False, interp=Image.BICUBIC):
# let the four corners of a rectangle to be ABCD, clockwise
if deg == 0:
return img
w, h = img.size
A = np.array([-w / 2, h / 2])
B = np.array([w / 2, h / 2])
C = np.array([w / 2, -h / 2])
D = np.array([-w / 2, -h / 2])
rad = np.radians(deg)
c, s = np.cos(rad), np.sin(rad)
R = np.array([[c, -s], [s, c]]).T
Arot = np.dot(A, R)
Brot = np.dot(B, R)
if deg > 0:
X = seg_intersect(A, C, Arot, Brot)
offset = X - A
offset[1] = -offset[1]
else:
X = seg_intersect(B, D, Arot, Brot)
offset = B - X
if same_size:
wh_org = np.array([w, h])
wh = np.ceil(np.divide(np.square(wh_org), wh_org - 2 * offset)).astype(np.int32)
offset = (wh - wh_org) / 2
img = img.resize(wh, interp)
w = wh[0]
h = wh[1]
else:
offset = np.ceil(offset)
img = img.rotate(deg, interp)
return img.crop(
(offset[0],
offset[1],
w - offset[0],
h - offset[1])
)