-
Notifications
You must be signed in to change notification settings - Fork 4
/
hlw_dataset.py
178 lines (135 loc) · 5.02 KB
/
hlw_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
import numpy as np
import torch
from torch.utils.data import Dataset
from torchvision import transforms
from skimage import io
from skimage import color
from skimage.draw import line, set_color, circle
import csv
import random
import math
class HLWDataset(Dataset):
"""
Access to the HLW dataset.
Horizon Lines in the Wild, version 1
http://www.cs.uky.edu/~jacobs/datasets/hlw/
Assumed to be in the ./hlw/ folder in the working directory.
"""
def __init__(self, data_file, imsize, training):
"""
Read image list and meta data.
data_file -- training/test split file
imsize -- rescale images to this max side length
training -- flag that toggles data augmentation
"""
self.imsize = imsize
self.training = training
# read image list
img_db = open(data_file, 'r')
self.images = ['hlw/images/' + f[:-1] for f in img_db.readlines()]
img_db.close()
# read ground truth labels
metadata_file = open('hlw/metadata.csv')
metadata = csv.reader(metadata_file)
self.gt = [None] * len(self.images)
for row in metadata:
cur_image = 'hlw/images/' + row[0]
if cur_image in self.images:
idx = self.images.index(cur_image)
cur_labels = torch.zeros((4))
cur_labels[0] = float(row[1])
cur_labels[1] = float(row[2])
cur_labels[2] = float(row[3])
cur_labels[3] = float(row[4])
self.gt[idx] = cur_labels
metadata_file.close()
def __len__(self):
return len(self.images)
def __getitem__(self, idx):
image = io.imread(self.images[idx])
label_scale = max(image.shape[0], image.shape[1]) # longer image side
image_scale = self.imsize / label_scale # normalize to longest side = 1
#convert ground truth coordinate system (from zero center to zero corner)
yOffset = image.shape[0] / label_scale / 2 # mind zero padding to make image square
xOffset = image.shape[1] / label_scale / 2
gt = self.gt[idx].clone()
gt /= label_scale
gt[0] += xOffset
gt[1] *= -1
gt[1] += yOffset
gt[2] += xOffset
gt[3] *= -1
gt[3] += yOffset
# convert image to RGB
if len(image.shape) < 3:
image = color.gray2rgb(image)
# original image dimensions
src_h = int(image.shape[0] * image_scale)
src_w = int(image.shape[1] * image_scale)
# resize and convert to gray scale
image = transforms.functional.to_pil_image(image)
image = transforms.functional.resize(image, (src_h, src_w))
image = transforms.functional.adjust_saturation(image, 0)
if self.training:
# data augmenation
random_shift = 8
random_shift_x = random.randint(-random_shift,random_shift)
random_shift_y = random.randint(-random_shift,random_shift)
random_angle = random.uniform(-5,5)
random_scale = random.uniform(0.8,1.2)
image = transforms.functional.adjust_contrast(image, random.uniform(0.8, 1.2))
image = transforms.functional.adjust_brightness(image, random.uniform(0.8, 1.2))
image = transforms.functional.affine(image, random_angle, (random_shift_x, random_shift_y), random_scale, 0, resample=2)
image = transforms.functional.to_tensor(image)
# image dimensions after augmentation
imh = image.size(1)
imw = image.size(2)
# zero pad image to make it square
padding_left = int((self.imsize - image.size(2)) / 2)
padding_right = self.imsize - image.size(2) - padding_left
padding_top = int((self.imsize - image.size(1)) / 2)
padding_bottom = self.imsize - image.size(1) - padding_top
padding = torch.nn.ZeroPad2d((padding_left, padding_right, padding_top, padding_bottom))
image = padding(image)
# normalization of color values (mean and stddev calculated offline over HLW training set)
img_mask = image.sum(0) > 0
image[:,img_mask] -= 0.45
image[:,img_mask] /= 0.25
# add padding offset due to augmentation to ground truth
gt[0] += padding_left / self.imsize
gt[1] += padding_top / self.imsize
gt[2] += padding_left / self.imsize
gt[3] += padding_top / self.imsize
if self.training:
# rotate and scale ground truth according to augmentation
a = random_angle * math.pi / 180
cos_a = math.cos(a)
sin_a = math.sin(a)
rot_off = 0.5
gt -= rot_off
r_x1 = cos_a * gt[0] - sin_a * gt[1]
r_y1 = sin_a * gt[0] + cos_a * gt[1]
r_x2 = cos_a * gt[2] - sin_a * gt[3]
r_y2 = sin_a * gt[2] + cos_a * gt[3]
gt[0] = r_x1
gt[1] = r_y1
gt[2] = r_x2
gt[3] = r_y2
gt *= random_scale
gt += rot_off
gt[0] += random_shift_x / self.imsize
gt[2] += random_shift_x / self.imsize
gt[1] += random_shift_y / self.imsize
gt[3] += random_shift_y / self.imsize
#calculate slope and intercept
labels = torch.zeros((2))
labels[1] = (gt[3] - gt[1]) / (gt[2] - gt[0])
labels[0] = gt[1] - labels[1] * gt[0]
# pre-compute start and end coordinate of line (used in loss)
xStart = padding_left / self.imsize
xEnd = xStart + imw / self.imsize
# meta data for calculating the HLW loss correctly:
# xStart -- at which x position does the GT line enter the image
# xEnd -- at which x position does the GT line leave the image
# imh -- what is the image height (without zero padding)
return image, labels, xStart, xEnd, imh, idx