File size: 9,793 Bytes
5e9bd47 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 |
import albumentations as A
from albumentations.augmentations.geometric.functional import safe_rotate_enlarged_img_size, _maybe_process_in_chunks, \
keypoint_rotate
import cv2
import math
import random
import numpy as np
def safe_rotate(
img: np.ndarray,
angle: int = 0,
interpolation: int = cv2.INTER_LINEAR,
value: int = None,
border_mode: int = cv2.BORDER_REFLECT_101,
):
old_rows, old_cols = img.shape[:2]
# getRotationMatrix2D needs coordinates in reverse order (width, height) compared to shape
image_center = (old_cols / 2, old_rows / 2)
# Rows and columns of the rotated image (not cropped)
new_rows, new_cols = safe_rotate_enlarged_img_size(angle=angle, rows=old_rows, cols=old_cols)
# Rotation Matrix
rotation_mat = cv2.getRotationMatrix2D(image_center, angle, 1.0)
# Shift the image to create padding
rotation_mat[0, 2] += new_cols / 2 - image_center[0]
rotation_mat[1, 2] += new_rows / 2 - image_center[1]
# CV2 Transformation function
warp_affine_fn = _maybe_process_in_chunks(
cv2.warpAffine,
M=rotation_mat,
dsize=(new_cols, new_rows),
flags=interpolation,
borderMode=border_mode,
borderValue=value,
)
# rotate image with the new bounds
rotated_img = warp_affine_fn(img)
return rotated_img
def keypoint_safe_rotate(keypoint, angle, rows, cols):
old_rows = rows
old_cols = cols
# Rows and columns of the rotated image (not cropped)
new_rows, new_cols = safe_rotate_enlarged_img_size(angle=angle, rows=old_rows, cols=old_cols)
col_diff = (new_cols - old_cols) / 2
row_diff = (new_rows - old_rows) / 2
# Shift keypoint
shifted_keypoint = (int(keypoint[0] + col_diff), int(keypoint[1] + row_diff), keypoint[2], keypoint[3])
# Rotate keypoint
rotated_keypoint = keypoint_rotate(shifted_keypoint, angle, rows=new_rows, cols=new_cols)
return rotated_keypoint
class SafeRotate(A.SafeRotate):
def __init__(
self,
limit=90,
interpolation=cv2.INTER_LINEAR,
border_mode=cv2.BORDER_REFLECT_101,
value=None,
mask_value=None,
always_apply=False,
p=0.5,
):
super(SafeRotate, self).__init__(
limit=limit,
interpolation=interpolation,
border_mode=border_mode,
value=value,
mask_value=mask_value,
always_apply=always_apply,
p=p)
def apply(self, img, angle=0, interpolation=cv2.INTER_LINEAR, **params):
return safe_rotate(
img=img, value=self.value, angle=angle, interpolation=interpolation, border_mode=self.border_mode)
def apply_to_keypoint(self, keypoint, angle=0, **params):
return keypoint_safe_rotate(keypoint, angle=angle, rows=params["rows"], cols=params["cols"])
class CropWhite(A.DualTransform):
def __init__(self, value=(255, 255, 255), pad=0, p=1.0):
super(CropWhite, self).__init__(p=p)
self.value = value
self.pad = pad
assert pad >= 0
def update_params(self, params, **kwargs):
super().update_params(params, **kwargs)
assert "image" in kwargs
img = kwargs["image"]
height, width, _ = img.shape
x = (img != self.value).sum(axis=2)
if x.sum() == 0:
return params
row_sum = x.sum(axis=1)
top = 0
while row_sum[top] == 0 and top+1 < height:
top += 1
bottom = height
while row_sum[bottom-1] == 0 and bottom-1 > top:
bottom -= 1
col_sum = x.sum(axis=0)
left = 0
while col_sum[left] == 0 and left+1 < width:
left += 1
right = width
while col_sum[right-1] == 0 and right-1 > left:
right -= 1
# crop_top = max(0, top - self.pad)
# crop_bottom = max(0, height - bottom - self.pad)
# crop_left = max(0, left - self.pad)
# crop_right = max(0, width - right - self.pad)
# params.update({"crop_top": crop_top, "crop_bottom": crop_bottom,
# "crop_left": crop_left, "crop_right": crop_right})
params.update({"crop_top": top, "crop_bottom": height - bottom,
"crop_left": left, "crop_right": width - right})
return params
def apply(self, img, crop_top=0, crop_bottom=0, crop_left=0, crop_right=0, **params):
height, width, _ = img.shape
img = img[crop_top:height - crop_bottom, crop_left:width - crop_right]
img = A.augmentations.pad_with_params(
img, self.pad, self.pad, self.pad, self.pad, border_mode=cv2.BORDER_CONSTANT, value=self.value)
return img
def apply_to_keypoint(self, keypoint, crop_top=0, crop_bottom=0, crop_left=0, crop_right=0, **params):
x, y, angle, scale = keypoint[:4]
return x - crop_left + self.pad, y - crop_top + self.pad, angle, scale
def get_transform_init_args_names(self):
return ('value', 'pad')
class PadWhite(A.DualTransform):
def __init__(self, pad_ratio=0.2, p=0.5, value=(255, 255, 255)):
super(PadWhite, self).__init__(p=p)
self.pad_ratio = pad_ratio
self.value = value
def update_params(self, params, **kwargs):
super().update_params(params, **kwargs)
assert "image" in kwargs
img = kwargs["image"]
height, width, _ = img.shape
side = random.randrange(4)
if side == 0:
params['pad_top'] = int(height * self.pad_ratio * random.random())
elif side == 1:
params['pad_bottom'] = int(height * self.pad_ratio * random.random())
elif side == 2:
params['pad_left'] = int(width * self.pad_ratio * random.random())
elif side == 3:
params['pad_right'] = int(width * self.pad_ratio * random.random())
return params
def apply(self, img, pad_top=0, pad_bottom=0, pad_left=0, pad_right=0, **params):
height, width, _ = img.shape
img = A.augmentations.pad_with_params(
img, pad_top, pad_bottom, pad_left, pad_right, border_mode=cv2.BORDER_CONSTANT, value=self.value)
return img
def apply_to_keypoint(self, keypoint, pad_top=0, pad_bottom=0, pad_left=0, pad_right=0, **params):
x, y, angle, scale = keypoint[:4]
return x + pad_left, y + pad_top, angle, scale
def get_transform_init_args_names(self):
return ('value', 'pad_ratio')
class SaltAndPepperNoise(A.DualTransform):
def __init__(self, num_dots, value=(0, 0, 0), p=0.5):
super().__init__(p)
self.num_dots = num_dots
self.value = value
def apply(self, img, **params):
height, width, _ = img.shape
num_dots = random.randrange(self.num_dots + 1)
for i in range(num_dots):
x = random.randrange(height)
y = random.randrange(width)
img[x, y] = self.value
return img
def apply_to_keypoint(self, keypoint, **params):
return keypoint
def get_transform_init_args_names(self):
return ('value', 'num_dots')
class ResizePad(A.DualTransform):
def __init__(self, height, width, interpolation=cv2.INTER_LINEAR, value=(255, 255, 255)):
super(ResizePad, self).__init__(always_apply=True)
self.height = height
self.width = width
self.interpolation = interpolation
self.value = value
def apply(self, img, interpolation=cv2.INTER_LINEAR, **params):
h, w, _ = img.shape
img = A.augmentations.geometric.functional.resize(
img,
height=min(h, self.height),
width=min(w, self.width),
interpolation=interpolation
)
h, w, _ = img.shape
pad_top = (self.height - h) // 2
pad_bottom = (self.height - h) - pad_top
pad_left = (self.width - w) // 2
pad_right = (self.width - w) - pad_left
img = A.augmentations.pad_with_params(
img,
pad_top,
pad_bottom,
pad_left,
pad_right,
border_mode=cv2.BORDER_CONSTANT,
value=self.value,
)
return img
def normalized_grid_distortion(
img,
num_steps=10,
xsteps=(),
ysteps=(),
*args,
**kwargs
):
height, width = img.shape[:2]
# compensate for smaller last steps in source image.
x_step = width // num_steps
last_x_step = min(width, ((num_steps + 1) * x_step)) - (num_steps * x_step)
xsteps[-1] *= last_x_step / x_step
y_step = height // num_steps
last_y_step = min(height, ((num_steps + 1) * y_step)) - (num_steps * y_step)
ysteps[-1] *= last_y_step / y_step
# now normalize such that distortion never leaves image bounds.
tx = width / math.floor(width / num_steps)
ty = height / math.floor(height / num_steps)
xsteps = np.array(xsteps) * (tx / np.sum(xsteps))
ysteps = np.array(ysteps) * (ty / np.sum(ysteps))
# do actual distortion.
return A.augmentations.functional.grid_distortion(img, num_steps, xsteps, ysteps, *args, **kwargs)
class NormalizedGridDistortion(A.augmentations.transforms.GridDistortion):
def apply(self, img, stepsx=(), stepsy=(), interpolation=cv2.INTER_LINEAR, **params):
return normalized_grid_distortion(img, self.num_steps, stepsx, stepsy, interpolation, self.border_mode,
self.value)
def apply_to_mask(self, img, stepsx=(), stepsy=(), **params):
return normalized_grid_distortion(
img, self.num_steps, stepsx, stepsy, cv2.INTER_NEAREST, self.border_mode, self.mask_value)
|