-
-
Notifications
You must be signed in to change notification settings - Fork 46k
/
mosaic_augmentation.py
186 lines (167 loc) · 6.9 KB
/
mosaic_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
176
177
178
179
180
181
182
183
184
185
186
"""Source: https://github.com/jason9075/opencv-mosaic-data-aug"""
import glob
import os
import random
from string import ascii_lowercase, digits
import cv2
import numpy as np
# Parameters
OUTPUT_SIZE = (720, 1280) # Height, Width
SCALE_RANGE = (0.4, 0.6) # if height or width lower than this scale, drop it.
FILTER_TINY_SCALE = 1 / 100
LABEL_DIR = ""
IMG_DIR = ""
OUTPUT_DIR = ""
NUMBER_IMAGES = 250
def main() -> None:
"""
Get images list and annotations list from input dir.
Update new images and annotations.
Save images and annotations in output dir.
"""
img_paths, annos = get_dataset(LABEL_DIR, IMG_DIR)
for index in range(NUMBER_IMAGES):
idxs = random.sample(range(len(annos)), 4)
new_image, new_annos, path = update_image_and_anno(
img_paths,
annos,
idxs,
OUTPUT_SIZE,
SCALE_RANGE,
filter_scale=FILTER_TINY_SCALE,
)
# Get random string code: '7b7ad245cdff75241935e4dd860f3bad'
letter_code = random_chars(32)
file_name = path.split(os.sep)[-1].rsplit(".", 1)[0]
file_root = f"{OUTPUT_DIR}/{file_name}_MOSAIC_{letter_code}"
cv2.imwrite(f"{file_root}.jpg", new_image, [cv2.IMWRITE_JPEG_QUALITY, 85])
print(f"Succeeded {index+1}/{NUMBER_IMAGES} with {file_name}")
annos_list = []
for anno in new_annos:
width = anno[3] - anno[1]
height = anno[4] - anno[2]
x_center = anno[1] + width / 2
y_center = anno[2] + height / 2
obj = f"{anno[0]} {x_center} {y_center} {width} {height}"
annos_list.append(obj)
with open(f"{file_root}.txt", "w") as outfile:
outfile.write("\n".join(line for line in annos_list))
def get_dataset(label_dir: str, img_dir: str) -> tuple[list, list]:
"""
- label_dir <type: str>: Path to label include annotation of images
- img_dir <type: str>: Path to folder contain images
Return <type: list>: List of images path and labels
"""
img_paths = []
labels = []
for label_file in glob.glob(os.path.join(label_dir, "*.txt")):
label_name = label_file.split(os.sep)[-1].rsplit(".", 1)[0]
with open(label_file) as in_file:
obj_lists = in_file.readlines()
img_path = os.path.join(img_dir, f"{label_name}.jpg")
boxes = []
for obj_list in obj_lists:
obj = obj_list.rstrip("\n").split(" ")
xmin = float(obj[1]) - float(obj[3]) / 2
ymin = float(obj[2]) - float(obj[4]) / 2
xmax = float(obj[1]) + float(obj[3]) / 2
ymax = float(obj[2]) + float(obj[4]) / 2
boxes.append([int(obj[0]), xmin, ymin, xmax, ymax])
if not boxes:
continue
img_paths.append(img_path)
labels.append(boxes)
return img_paths, labels
def update_image_and_anno(
all_img_list: list,
all_annos: list,
idxs: list[int],
output_size: tuple[int, int],
scale_range: tuple[float, float],
filter_scale: float = 0.0,
) -> tuple[list, list, str]:
"""
- all_img_list <type: list>: list of all images
- all_annos <type: list>: list of all annotations of specific image
- idxs <type: list>: index of image in list
- output_size <type: tuple>: size of output image (Height, Width)
- scale_range <type: tuple>: range of scale image
- filter_scale <type: float>: the condition of downscale image and bounding box
Return:
- output_img <type: narray>: image after resize
- new_anno <type: list>: list of new annotation after scale
- path[0] <type: string>: get the name of image file
"""
output_img = np.zeros([output_size[0], output_size[1], 3], dtype=np.uint8)
scale_x = scale_range[0] + random.random() * (scale_range[1] - scale_range[0])
scale_y = scale_range[0] + random.random() * (scale_range[1] - scale_range[0])
divid_point_x = int(scale_x * output_size[1])
divid_point_y = int(scale_y * output_size[0])
new_anno = []
path_list = []
for i, index in enumerate(idxs):
path = all_img_list[index]
path_list.append(path)
img_annos = all_annos[index]
img = cv2.imread(path)
if i == 0: # top-left
img = cv2.resize(img, (divid_point_x, divid_point_y))
output_img[:divid_point_y, :divid_point_x, :] = img
for bbox in img_annos:
xmin = bbox[1] * scale_x
ymin = bbox[2] * scale_y
xmax = bbox[3] * scale_x
ymax = bbox[4] * scale_y
new_anno.append([bbox[0], xmin, ymin, xmax, ymax])
elif i == 1: # top-right
img = cv2.resize(img, (output_size[1] - divid_point_x, divid_point_y))
output_img[:divid_point_y, divid_point_x : output_size[1], :] = img
for bbox in img_annos:
xmin = scale_x + bbox[1] * (1 - scale_x)
ymin = bbox[2] * scale_y
xmax = scale_x + bbox[3] * (1 - scale_x)
ymax = bbox[4] * scale_y
new_anno.append([bbox[0], xmin, ymin, xmax, ymax])
elif i == 2: # bottom-left
img = cv2.resize(img, (divid_point_x, output_size[0] - divid_point_y))
output_img[divid_point_y : output_size[0], :divid_point_x, :] = img
for bbox in img_annos:
xmin = bbox[1] * scale_x
ymin = scale_y + bbox[2] * (1 - scale_y)
xmax = bbox[3] * scale_x
ymax = scale_y + bbox[4] * (1 - scale_y)
new_anno.append([bbox[0], xmin, ymin, xmax, ymax])
else: # bottom-right
img = cv2.resize(
img, (output_size[1] - divid_point_x, output_size[0] - divid_point_y)
)
output_img[
divid_point_y : output_size[0], divid_point_x : output_size[1], :
] = img
for bbox in img_annos:
xmin = scale_x + bbox[1] * (1 - scale_x)
ymin = scale_y + bbox[2] * (1 - scale_y)
xmax = scale_x + bbox[3] * (1 - scale_x)
ymax = scale_y + bbox[4] * (1 - scale_y)
new_anno.append([bbox[0], xmin, ymin, xmax, ymax])
# Remove bounding box small than scale of filter
if filter_scale > 0:
new_anno = [
anno
for anno in new_anno
if filter_scale < (anno[3] - anno[1]) and filter_scale < (anno[4] - anno[2])
]
return output_img, new_anno, path_list[0]
def random_chars(number_char: int) -> str:
"""
Automatic generate random 32 characters.
Get random string code: '7b7ad245cdff75241935e4dd860f3bad'
>>> len(random_chars(32))
32
"""
assert number_char > 1, "The number of character should greater than 1"
letter_code = ascii_lowercase + digits
return "".join(random.choice(letter_code) for _ in range(number_char))
if __name__ == "__main__":
main()
print("DONE ✅")