-
Notifications
You must be signed in to change notification settings - Fork 34
/
road_dataset.py
369 lines (300 loc) · 12 KB
/
road_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
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
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
import collections
import math
import os
import random
import cv2
import numpy as np
import torch
from data_utils import affinity_utils
from torch.utils import data
class RoadDataset(data.Dataset):
def __init__(
self, config, dataset_name, seed=7, multi_scale_pred=True, is_train=True
):
# Seed
np.random.seed(seed)
torch.manual_seed(seed)
random.seed(seed)
self.split = "train" if is_train else "val"
self.config = config
# paths
self.dir = self.config[dataset_name]["dir"]
self.img_root = os.path.join(self.dir, "images/")
self.gt_root = os.path.join(self.dir, "gt/")
self.image_list = self.config[dataset_name]["file"]
# list of all images
self.images = [line.rstrip("\n") for line in open(self.image_list)]
# augmentations
self.augmentation = self.config["augmentation"]
self.crop_size = [
self.config[dataset_name]["crop_size"],
self.config[dataset_name]["crop_size"],
]
self.multi_scale_pred = multi_scale_pred
# preprocess
self.angle_theta = self.config["angle_theta"]
self.mean_bgr = np.array(eval(self.config["mean"]))
self.deviation_bgr = np.array(eval(self.config["std"]))
self.normalize_type = self.config["normalize_type"]
# to avoid Deadloack between CV Threads and Pytorch Threads caused in resizing
cv2.setNumThreads(0)
self.files = collections.defaultdict(list)
for f in self.images:
self.files[self.split].append(
{
"img": self.img_root
+ f
+ self.config[dataset_name]["image_suffix"],
"lbl": self.gt_root + f + self.config[dataset_name]["gt_suffix"],
}
)
def __len__(self):
return len(self.files[self.split])
def getRoadData(self, index):
image_dict = self.files[self.split][index]
# read each image in list
if os.path.isfile(image_dict["img"]):
image = cv2.imread(image_dict["img"]).astype(np.float)
else:
print("ERROR: couldn't find image -> ", image_dict["img"])
if os.path.isfile(image_dict["lbl"]):
gt = cv2.imread(image_dict["lbl"], 0).astype(np.float)
else:
print("ERROR: couldn't find image -> ", image_dict["lbl"])
if self.split == "train":
image, gt = self.random_crop(image, gt, self.crop_size)
else:
image = cv2.resize(
image,
(self.crop_size[0], self.crop_size[1]),
interpolation=cv2.INTER_LINEAR,
)
gt = cv2.resize(
gt,
(self.crop_size[0], self.crop_size[1]),
interpolation=cv2.INTER_LINEAR,
)
if self.split == "train" and index == len(self.files[self.split]) - 1:
np.random.shuffle(self.files[self.split])
h, w, c = image.shape
if self.augmentation == 1:
flip = np.random.choice(2) * 2 - 1
image = np.ascontiguousarray(image[:, ::flip, :])
gt = np.ascontiguousarray(gt[:, ::flip])
rotation = np.random.randint(4) * 90
M = cv2.getRotationMatrix2D((w / 2, h / 2), rotation, 1)
image = cv2.warpAffine(image, M, (w, h))
gt = cv2.warpAffine(gt, M, (w, h))
image = self.reshape(image)
image = torch.from_numpy(np.array(image))
return image, gt
def getOrientationGT(self, keypoints, height, width):
vecmap, vecmap_angles = affinity_utils.getVectorMapsAngles(
(height, width), keypoints, theta=self.angle_theta, bin_size=10
)
vecmap_angles = torch.from_numpy(vecmap_angles)
return vecmap_angles
def getCorruptRoad(
self, road_gt, height, width, artifacts_shape="linear", element_counts=8
):
# False Negative Mask
FNmask = np.ones((height, width), np.float)
# False Positive Mask
FPmask = np.zeros((height, width), np.float)
indices = np.where(road_gt == 1)
if artifacts_shape == "square":
shapes = [[16, 16], [32, 32]]
##### FNmask
if len(indices[0]) == 0: ### no road pixel in GT
pass
else:
for c_ in range(element_counts):
c = np.random.choice(len(shapes), 1)[
0
] ### choose random square size
shape_ = shapes[c]
ind = np.random.choice(len(indices[0]), 1)[
0
] ### choose a random road pixel as center for the square
row = indices[0][ind]
col = indices[1][ind]
FNmask[
row - shape_[0] / 2 : row + shape_[0] / 2,
col - shape_[1] / 2 : col + shape_[1] / 2,
] = 0
#### FPmask
for c_ in range(element_counts):
c = np.random.choice(len(shapes), 2)[0] ### choose random square size
shape_ = shapes[c]
row = np.random.choice(height - shape_[0] - 1, 1)[
0
] ### choose random pixel
col = np.random.choice(width - shape_[1] - 1, 1)[
0
] ### choose random pixel
FPmask[
row - shape_[0] / 2 : row + shape_[0] / 2,
col - shape_[1] / 2 : col + shape_[1] / 2,
] = 1
elif artifacts_shape == "linear":
##### FNmask
if len(indices[0]) == 0: ### no road pixel in GT
pass
else:
for c_ in range(element_counts):
c1 = np.random.choice(len(indices[0]), 1)[
0
] ### choose random 2 road pixels to draw a line
c2 = np.random.choice(len(indices[0]), 1)[0]
cv2.line(
FNmask,
(indices[1][c1], indices[0][c1]),
(indices[1][c2], indices[0][c2]),
0,
self.angle_theta * 2,
)
#### FPmask
for c_ in range(element_counts):
row1 = np.random.choice(height, 1)
col1 = np.random.choice(width, 1)
row2, col2 = (
row1 + np.random.choice(50, 1),
col1 + np.random.choice(50, 1),
)
cv2.line(FPmask, (col1, row1), (col2, row2), 1, self.angle_theta * 2)
erased_gt = (road_gt * FNmask) + FPmask
erased_gt[erased_gt > 0] = 1
return erased_gt
def reshape(self, image):
if self.normalize_type == "Std":
image = (image - self.mean_bgr) / (3 * self.deviation_bgr)
elif self.normalize_type == "MinMax":
image = (image - self.min_bgr) / (self.max_bgr - self.min_bgr)
image = image * 2 - 1
elif self.normalize_type == "Mean":
image -= self.mean_bgr
else:
image = (image / 255.0) * 2 - 1
image = image.transpose(2, 0, 1)
return image
def random_crop(self, image, gt, size):
w, h, _ = image.shape
crop_h, crop_w = size
start_x = np.random.randint(0, w - crop_w)
start_y = np.random.randint(0, h - crop_h)
image = image[start_x : start_x + crop_w, start_y : start_y + crop_h, :]
gt = gt[start_x : start_x + crop_w, start_y : start_y + crop_h]
return image, gt
class SpacenetDataset(RoadDataset):
def __init__(self, config, seed=7, multi_scale_pred=True, is_train=True):
super(SpacenetDataset, self).__init__(
config, "spacenet", seed, multi_scale_pred, is_train
)
# preprocess
self.threshold = self.config["thresh"]
print("Threshold is set to {} for {}".format(self.threshold, self.split))
def __getitem__(self, index):
image, gt = self.getRoadData(index)
c, h, w = image.shape
labels = []
vecmap_angles = []
if self.multi_scale_pred:
smoothness = [1, 2, 4]
scale = [4, 2, 1]
else:
smoothness = [4]
scale = [1]
for i, val in enumerate(scale):
if val != 1:
gt_ = cv2.resize(
gt,
(int(math.ceil(h / (val * 1.0))), int(math.ceil(w / (val * 1.0)))),
interpolation=cv2.INTER_NEAREST,
)
else:
gt_ = gt
gt_orig = np.copy(gt_)
gt_orig /= 255.0
gt_orig[gt_orig < self.threshold] = 0
gt_orig[gt_orig >= self.threshold] = 1
labels.append(gt_orig)
keypoints = affinity_utils.getKeypoints(
gt_, thresh=0.98, smooth_dist=smoothness[i]
)
vecmap_angle = self.getOrientationGT(
keypoints,
height=int(math.ceil(h / (val * 1.0))),
width=int(math.ceil(w / (val * 1.0))),
)
vecmap_angles.append(vecmap_angle)
return image, labels, vecmap_angles
class DeepGlobeDataset(RoadDataset):
def __init__(self, config, seed=7, multi_scale_pred=True, is_train=True):
super(DeepGlobeDataset, self).__init__(
config, "deepglobe", seed, multi_scale_pred, is_train
)
pass
def __getitem__(self, index):
image, gt = self.getRoadData(index)
c, h, w = image.shape
labels = []
vecmap_angles = []
if self.multi_scale_pred:
smoothness = [1, 2, 4]
scale = [4, 2, 1]
else:
smoothness = [4]
scale = [1]
for i, val in enumerate(scale):
if val != 1:
gt_ = cv2.resize(
gt,
(int(math.ceil(h / (val * 1.0))), int(math.ceil(w / (val * 1.0)))),
interpolation=cv2.INTER_NEAREST,
)
else:
gt_ = gt
gt_orig = np.copy(gt_)
gt_orig /= 255.0
labels.append(gt_orig)
# Create Orientation Ground Truth
keypoints = affinity_utils.getKeypoints(
gt_orig, is_gaussian=False, smooth_dist=smoothness[i]
)
vecmap_angle = self.getOrientationGT(
keypoints,
height=int(math.ceil(h / (val * 1.0))),
width=int(math.ceil(w / (val * 1.0))),
)
vecmap_angles.append(vecmap_angle)
return image, labels, vecmap_angles
class SpacenetDatasetCorrupt(RoadDataset):
def __init__(self, config, seed=7, is_train=True):
super(SpacenetDatasetCorrupt, self).__init__(
config, "spacenet", seed, multi_scale_pred=False, is_train=is_train
)
# preprocess
self.threshold = self.config["thresh"]
print("Threshold is set to {} for {}".format(self.threshold, self.split))
def __getitem__(self, index):
image, gt = self.getRoadData(index)
c, h, w = image.shape
gt /= 255.0
gt[gt < self.threshold] = 0
gt[gt >= self.threshold] = 1
erased_gt = self.getCorruptRoad(gt.copy(), h, w)
erased_gt = torch.from_numpy(erased_gt)
return image, [gt], [erased_gt]
class DeepGlobeDatasetCorrupt(RoadDataset):
def __init__(self, config, seed=7, is_train=True):
super(DeepGlobeDatasetCorrupt, self).__init__(
config, "deepglobe", seed, multi_scale_pred=False, is_train=is_train
)
pass
def __getitem__(self, index):
image, gt = self.getRoadData(index)
c, h, w = image.shape
gt /= 255.0
erased_gt = self.getCorruptRoad(gt, h, w)
erased_gt = torch.from_numpy(erased_gt)
return image, [gt], [erased_gt]