forked from jonathan-hellwig/robot_detection
-
Notifications
You must be signed in to change notification settings - Fork 0
/
datasets.py
309 lines (273 loc) · 11.1 KB
/
datasets.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
import os
import shutil
import torch
import numpy as np
import tqdm
from PIL import Image, ImageDraw
import torchvision
import torchvision.transforms as T
import albumentations as A
import utils
class RoboEireanData(torch.utils.data.Dataset):
"""Custom PyTorch Dataset class for loading RoboEireann dataset.
Args:
data_path (str): Path to dataset directory.
selected_classes (List[str]): List of classes to select from ["ball", "robot", "goal_post", "penalty_spot"].
image_transforms (optional): Torchvision transforms to apply on the images. Defaults to None.
bounding_box_transforms (optional): A function that takes in a tensor of bounding boxes and applies the transformation. Defaults to None.
"""
CLASSES = ["ball", "robot", "goal_post", "penalty_spot"]
def __init__(
self,
data_path: str,
selected_classes: list[str],
image_transforms=None,
bounding_box_transforms=None,
) -> None:
assert set(selected_classes) <= set(self.CLASSES)
self.image_transforms = image_transforms
self.bounding_box_transforms = bounding_box_transforms
self.data_path = data_path
self.selected_classes = selected_classes
self.images = sorted(os.listdir(os.path.join(data_path, "images")))
self.labels = sorted(os.listdir(os.path.join(data_path, "labels")))
def __getitem__(self, idx):
assert self.images[idx][:-3] == self.labels[idx][:-3]
image_path = os.path.join(self.data_path, "images", self.images[idx])
label_path = os.path.join(self.data_path, "labels", self.labels[idx])
image = Image.open(image_path)
label_strings = open(label_path).read().splitlines()
target_bounding_boxes = []
target_classes = []
for label_string in label_strings:
parsed_target_class = int(label_string[0])
if self.CLASSES[parsed_target_class] in self.selected_classes:
target_classes.append(
self.selected_classes.index(self.CLASSES[parsed_target_class])
)
target_bounding_boxes.append(
np.fromstring(label_string[1:], sep=" ", dtype=np.float32)
)
target_bounding_boxes = torch.tensor(np.array(target_bounding_boxes))
target_classes = torch.tensor(np.array(target_classes)) + 1
target_classes = target_classes.unsqueeze(1)
if self.bounding_box_transforms:
target_bounding_boxes = self.bounding_box_transforms(target_bounding_boxes)
if self.image_transforms:
image = self.image_transforms(image)
return image, target_bounding_boxes, target_classes, image_path
def __len__(self):
return len(self.images)
class RoboEireanDataWithEncoder(torch.utils.data.Dataset):
"""
A PyTorch Dataset for the RoboEirean dataset that encodes bounding boxes using an Encoder object.
Args:
data_path (str): The path to the RoboEirean dataset.
encoder (utils.Encoder): An Encoder object that will be used to encode the bounding boxes.
selected_classes (list[str]): A list of the classes to include in the dataset.
image_transforms (optional): A transformation to apply to the images.
bounding_box_transforms (optional): A transformation to apply to the bounding boxes.
"""
def __init__(
self,
data_path: str,
encoder: utils.Encoder,
selected_classes: list[str],
image_transforms=None,
bounding_box_transforms=None,
) -> None:
self.dataset = RoboEireanData(
data_path, selected_classes, image_transforms, bounding_box_transforms,
)
self.encoder = encoder
def __getitem__(self, idx):
image, target_bounding_boxes, target_classes, image_path = self.dataset[idx]
(
encoded_bounding_boxes,
target_mask,
target_classes,
) = self.encoder.apply(target_bounding_boxes, target_classes)
return image, encoded_bounding_boxes, target_mask, target_classes, image_path
def __len__(self):
return len(self.dataset)
class TransformedRoboEireanData(torch.utils.data.Dataset):
def __init__(
self,
data_path: str,
encoder: utils.Encoder,
) -> None:
self.encoder = encoder
loaded_images = torch.load(os.path.join(data_path, "transformed_images.pt"))
bounding_boxes = torch.load(os.path.join(data_path, "target_bounding_boxes.pt"))
object_classes = torch.load(os.path.join(data_path, "target_classes.pt"))
self.images = []
self.encoded_bounding_boxes = []
self.encoded_target_classes = []
self.target_masks = []
for bounding_box, object_class, image in zip(
bounding_boxes, object_classes, loaded_images
):
(
encoded_bounding_boxes,
target_mask,
target_classes,
) = self.encoder.apply(bounding_box, object_class)
self.encoded_bounding_boxes.append(encoded_bounding_boxes)
self.encoded_target_classes.append(target_classes)
self.target_masks.append(target_mask)
self.images.append(image)
def __getitem__(self, idx):
return (
self.images[idx],
self.encoded_bounding_boxes[idx],
self.target_masks[idx],
self.encoded_target_classes[idx],
)
def __len__(self):
return len(self.encoded_bounding_boxes)
class SyntheticData(torch.utils.data.Dataset):
"""
SyntheticData generates a synthetic dataset of images with bounding boxes and classes.
Args:
image_width (int): Width of the generated images.
image_height (int): Height of the generated images.
length (int): Number of images in the dataset.
encoder (utils.Encoder): An encoder object to encode bounding boxes and classes.
"""
def __init__(
self, image_width: int, image_height: int, length: int, encoder: utils.Encoder
):
self.image_width = image_width
self.image_height = image_height
self.length = length
self.encoder = encoder
self.images = []
self.encoded_bounding_boxes = []
self.encoded_target_classes = []
self.target_masks = []
for _ in range(length):
image, bounding_box = self._generate_image()
encoded_bounding_boxes, target_mask, target_classes = self.encoder.apply(
bounding_box, torch.tensor([[1]])
)
self.images.append(image)
self.encoded_bounding_boxes.append(encoded_bounding_boxes)
self.encoded_target_classes.append(target_classes)
self.target_masks.append(target_mask)
def __getitem__(self, idx: int):
return (
self.images[idx],
self.encoded_bounding_boxes[idx],
self.target_masks[idx],
self.encoded_target_classes[idx],
)
def __len__(self) -> int:
return self.length
def _generate_image(self):
image = Image.new("1", (self.image_width, self.image_height))
image_draw = ImageDraw.Draw(image)
center = torch.tensor(
[
torch.randint(0, self.image_width - 1, (1,)).item(),
torch.randint(0, self.image_height - 1, (1,)).item(),
]
)
size = torch.tensor([self.image_width * 0.25, self.image_height * 0.25])
upper_left = center - size / 2
lower_right = center + size / 2
image_draw.rectangle(
[upper_left[0], upper_left[1], lower_right[0], lower_right[1]], fill=255
)
bounding_box = torch.tensor(
[
[
center[0].item() / self.image_width,
center[1].item() / self.image_height,
size[0].item() / self.image_width,
size[1].item() / self.image_height,
]
]
)
return (
torchvision.transforms.functional.pil_to_tensor(image).to(torch.float),
bounding_box,
)
def calculate_mean_std(dataset: RoboEireanData):
means = []
for image, _, _ in tqdm.tqdm(dataset):
means.append(torch.mean(image, dim=(1, 2)))
stacked_means = torch.stack(means)
mean = torch.mean(stacked_means, dim=0)
std = torch.std(stacked_means, dim=0)
return mean, std
def preprocess_data(
base_path: str,
split_path: str,
image_augmentations=[lambda x: x],
bounding_box_augmentations=[lambda x: x],
):
image_transforms = T.Compose(
[
T.Grayscale(),
T.PILToTensor(),
T.ConvertImageDtype(torch.float32),
T.Resize((60, 80)),
]
)
bounding_box_transforms = T.Compose([])
train_data = RoboEireanData(
os.path.join(base_path, "raw", split_path),
["robot"],
image_transforms=image_transforms,
bounding_box_transforms=bounding_box_transforms,
)
images = []
image_means = []
image_stds = []
target_bounding_boxes = []
target_classes = []
for image, target_bounding_box, target_class in tqdm.tqdm(train_data):
for image_augmentation, bounding_box_augmentation in zip(
image_augmentations, bounding_box_augmentations
):
images.append(image_augmentation(image))
target_bounding_boxes.append(bounding_box_augmentation(target_bounding_box))
target_classes.append(target_class)
image_means.append(torch.mean(images[-1], dim=(1, 2)))
image_stds.append(torch.std(images[-1], dim=(1, 2)))
image_means = torch.tensor(image_means)
image_stds = torch.tensor(image_stds)
image_mean = torch.mean(image_means)
image_std = torch.mean(image_stds)
stacked_images = (torch.stack(images) - image_mean) / image_std
transformed_images_path = os.path.join(
base_path, "transformed", split_path, "transformed_images.pt"
)
target_bounding_boxes_path = os.path.join(
base_path, "transformed", split_path, "target_bounding_boxes.pt"
)
target_classes_path = os.path.join(
base_path, "transformed", split_path, "target_classes.pt"
)
image_normalize_path = os.path.join(
base_path, "transformed", split_path, "image_normalize.pt"
)
torch.save(stacked_images, transformed_images_path)
torch.save(target_bounding_boxes, target_bounding_boxes_path)
torch.save(target_classes, target_classes_path)
torch.save(torch.tensor([image_mean, image_std]), image_normalize_path)
def flip_bounding_boxes(bounding_boxes: torch.Tensor) -> torch.Tensor:
bounding_boxes = bounding_boxes.clone()
if bounding_boxes.dim() == 2:
bounding_boxes[:, 0] = 1 - bounding_boxes[:, 0]
return bounding_boxes
if __name__ == "__main__":
image_augmentations = [lambda x: x, torchvision.transforms.functional.hflip]
bounding_box_augmentations = [lambda x: x, flip_bounding_boxes]
preprocess_data("data", "val")
preprocess_data(
"data",
"train",
image_augmentations,
bounding_box_augmentations,
)