-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataset.py
217 lines (168 loc) · 6.73 KB
/
dataset.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
"""
Filename [ dataset.py ]
PackageName [ DLCV Spring 2019 - YOLOv1 ]
Synposis [ DataLoader of the aerial dataset ]
"""
import os
import random
import sys
import time
import numpy as np
import torch
import torch.utils.data as data
import torchvision.transforms as transforms
from PIL import Image, ImageEnhance, ImageFilter
from sklearn.preprocessing import LabelEncoder, OneHotEncoder
from torch.utils.data import DataLoader, Dataset
import utils
__all__ = ['MyDataset']
class MyDataset(Dataset):
classnames = utils.classnames
labelEncoder = utils.labelEncoder
oneHotEncoder = utils.oneHotEncoder
def __init__(self, root, grid_num=7, bbox_num=2, class_num=16, train=True, transform=None):
""" Save the imageNames and the labelNames. """
self.filenames = []
self.train = train
self.transform = transform
self.grid_num = grid_num
self.bbox_num = bbox_num
self.class_num = class_num
image_folder = os.path.join(root, "images")
anno_folder = os.path.join(root, "labelTxt_hbb")
imageNames = os.listdir(image_folder)
for name in imageNames:
imageName = os.path.join(image_folder, name)
labelName = os.path.join(anno_folder, name.split(".")[0] + ".txt")
self.filenames.append((imageName, labelName))
def __len__(self):
return len(self.filenames)
def __getitem__(self, index):
imageName, labelName = self.filenames[index]
image = Image.open(imageName)
boxes, classIndexs = self.readtxt(labelName)
if self.train:
image = self.RandomAdjustHSV(image, 0.95, 1.05)
if random.random() < 0.5: image, boxes = self.HorizontalFlip(image, boxes)
if random.random() < 0.5: image, boxes = self.VerticalFlip(image, boxes)
target = self.encoder(boxes, classIndexs, image.size)
target = torch.from_numpy(target)
if self.transform:
image = self.transform(image)
return image, target, labelName
def encoder(self, boxes, classindex, image_size):
"""
Parameters
----------
boxes : numpy.array
[N, 4], contains [x1, y1, x2, y2] in integers
labels : numpy.array
[N, self.class_num]
Return
------
targets : numpy.array
[self.grid_num, self.grid_num, self.class_num]
"""
image_size = np.asarray(image_size)
image_size = np.concatenate((image_size, image_size), axis=0)
target = np.zeros((self.grid_num, self.grid_num, 5 * self.bbox_num + self.class_num))
boxes = boxes / image_size
cell_size = 1. / self.grid_num
wh = boxes[:, 2:] - boxes[:, :2]
centerXY = (boxes[:, 2:] + boxes[:, :2]) / 2
ij = (np.ceil(centerXY / cell_size) - 1).astype(int)
# Confidence
for index, (i, j) in enumerate(ij):
# print("Index: {}, i: {}, j: {}".format(index, i, j))
target[j, i] = 0 # Reset as zero
target[j, i, 4] = 1
target[j, i, 9] = 1
target[j, i, classindex + 10] = 1
# Coordinate transform to xyhw
cornerXY = ij[index] * cell_size
deltaXY = (centerXY[index] - cornerXY) / cell_size
target[j, i, 2:4] = wh[index]
target[j, i, :2] = deltaXY
target[j, i, 7:9] = wh[index]
target[j, i, 5:7] = deltaXY
# Target in numpy
return target
def readtxt(self, labelName):
"""
Transfer the labels to the tensor.
Parameters
----------
labelName: str
the label textfile to open
Return
------
target: np.array
[7 * 7 * 26]
"""
with open(labelName, "r") as textfile:
labels = textfile.readlines()
labels = np.asarray("".join(labels).replace("\n", " ").strip().split()).reshape(-1, 10)
classNames = np.asarray(labels[:, 8])
classIndexs = self.labelEncoder.transform(classNames)
boxes = np.asarray(labels[:, :8]).astype(np.float)
boxes = np.concatenate((boxes[:, :2], boxes[:, 4:6]), axis=1)
return boxes, classIndexs
def RandomAdjustHSV(self, img, min_f, max_f, prob=0.5):
""" Augmentation Method: Adjust HSV """
if random.random() < prob:
factor = random.uniform(min_f, max_f)
choice = random.randint(0, 3)
if choice == 0:
img = ImageEnhance.Color(img).enhance(factor)
elif choice == 1:
img = ImageEnhance.Brightness(img).enhance(factor)
elif choice == 2:
img = ImageEnhance.Contrast(img).enhance(factor)
elif choice == 3:
img = ImageEnhance.Sharpness(img).enhance(factor)
return img
def HorizontalFlip(self, im, boxes):
""" Augmentation Method: Horizontal Flip """
im = im.transpose(Image.FLIP_LEFT_RIGHT)
h, w = im.size
xmin = w - boxes[:, 2]
xmax = w - boxes[:, 0]
boxes[:, 0] = xmin
boxes[:, 2] = xmax
return im, boxes
def VerticalFlip(self, im, boxes):
""" Augmentation Method: Vertical Flip """
im = im.transpose(Image.FLIP_TOP_BOTTOM)
h, w = im.size
ymin = h - boxes[:, 3]
ymax = h - boxes[:, 1]
boxes[:, 1] = ymin
boxes[:, 3] = ymax
return im, boxes
def ZoomIn(self, im, boxes, scale):
""" Augmentation Method: Zoom In (Not Suggest) """
h, w = im.size
boundary = int(w * (scale - 1) / 2)
im = im.resize((int(h * scale), int(w * scale)), Image.ANTIALIAS)
im = im.crop((boundary, boundary, boundary + h, boundary + w))
boxes = (boxes * scale - boundary).astype(int).clip(min=0, max=w)
return im, boxes
class Testset(Dataset):
def __init__(self, img_root, grid_num=7, bbox_num=2, class_num=16, transform=None):
""" Save the imageNames and the labelNames and read in future. """
self.filenames = [ os.path.join(img_root, name) for name in os.listdir(img_root) ]
self.transform = transform
self.grid_num = grid_num
self.bbox_num = bbox_num
self.class_num = class_num
def __len__(self):
return len(self.filenames)
def __getitem__(self, index):
image = Image.open(self.filenames[index])
if self.transform:
image = self.transform(image)
return image, self.filenames[index].split(".")[0]
def main():
return
if __name__ == "__main__":
main()