forked from cmusatyalab/change-detection
-
Notifications
You must be signed in to change notification settings - Fork 0
/
vl_cmu_cd_dataset.py
226 lines (194 loc) · 8.67 KB
/
vl_cmu_cd_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
import torch
from torch.utils.data import Dataset
from torchvision import transforms, datasets
from PIL import Image
import numpy as np
from skimage import io
import argparse
import matplotlib.pyplot as plt
import pickle
import time
import h5py
import glob
import os
import time
import numpy.random
import pdb
#from pdb import set_trace as st
MaskOut = [0,0,0]
NoChange = [255, 255, 255]
Barrier = [136, 0, 21]
Bin = [237, 28, 36]
ConstructionMaintenance = [255, 127, 39]
Misc = [255, 242, 0] # this class is a mistake, only one train example, and it's actually Rubbish class...
OtherObjects = [34, 177, 76]
PersonCycle = [0, 162, 232]
Rubbish = [63, 72, 204]
Sign = [163, 73, 164]
TrafficCone = [255, 174, 201]
Vehicle = [181, 230, 29]
CLASS_DICT = np.array([NoChange, Barrier, Bin, ConstructionMaintenance,
OtherObjects, PersonCycle, Rubbish, Sign, TrafficCone, Vehicle])
CHANGE_CLASSES = ["NoChange", "Barrier", "Bin", "ConstructionMaintenance",
"OtherObjects", "PersonCycle", "Rubbish", "Sign", "TrafficCone", "Vehicle"]
''' Returns 224 x 224 x 11 vector '''
def oldprepareGTData(mask, num_class = 10):
new_mask = np.zeros(mask.shape[0:2] + (num_class,), dtype=np.float32)
# Mask Out Class
new_mask[np.all(mask == np.zeros((1,1,3), dtype=np.int32) , axis=2),0] = 1
# All other classes
for i in range(num_class):
new_mask[np.all(mask == CLASS_DICT[i].reshape(1,1,3) , axis=2),i] = 1
return new_mask
''' Returns 224 x 224 x 1 vector '''
def prepareGTData(mask, num_class = 10):
#pdb.set_trace()
mask = mask.astype('int')
new_mask = np.zeros(mask.shape[0:2], dtype=np.int64)
# Mask Out Class
new_mask[np.all(mask == np.zeros((1,1,3), dtype=np.int32) , axis=2)] = np.int64(-1)
new_mask[np.all(mask == np.array(Misc, dtype=np.int32).reshape(1,1,3) , axis=2)] = np.int64(4)
#OtherObjects
# All other classes
for i in range(num_class):
new_mask[np.all(mask == CLASS_DICT[i].reshape(1,1,3) , axis=2)] = np.int64(i)
return new_mask
class VLCMUCDDataset(Dataset):
''''
Dataset Class for VL-CMU-CD Dataset
'''
def __init__(self, file_paths, training=False, probability=0.5):
with open(file_paths, "rb") as fp:
self.vlcmu_data = pickle.load(fp)
self.resize_image = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize((224,224)),
])
self.max_rot = 15
self.max_trans = 0.2
self.max_scale_diff = 0.2
self.max_shear_deg = 10
self.training = training
self.p = probability
self.normalize_transform = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
def __len__(self):
return len(self.vlcmu_data['gt'])
def __getitem__(self, idx):
im1 = io.imread(self.vlcmu_data['im1'][idx])
im2 = io.imread(self.vlcmu_data['im2'][idx])
gt = io.imread(self.vlcmu_data['gt'][idx])
# print(self.vlcmu_data['im1'][idx])
# print(self.vlcmu_data['im2'][idx])
# print(self.vlcmu_data['gt'][idx])
im1 = self.resize_image(im1).convert('RGB')
im2 = self.resize_image(im2).convert('RGB')
gt = self.resize_image(gt).convert('RGB')
if self.training:
# Shear
if (np.random.rand() < self.p):
angle = int( self.max_rot * np.random.rand() )
translation = (self.max_trans * np.random.rand() , self.max_trans * np.random.rand() )
scale = self.max_scale_diff * 2. * (np.random.rand() - 0.5) + 1.
shear = int( self.max_shear_deg * np.random.rand() )
im1 = transforms.functional.affine(im1, angle=angle, translate=translation,
scale=scale, shear=shear, resample=Image.NEAREST, fillcolor=None)
im2 = transforms.functional.affine(im2, angle=angle, translate=translation,
scale=scale, shear=shear, resample=Image.NEAREST, fillcolor=None)
gt = transforms.functional.affine(gt, angle=angle, translate=translation,
scale=scale, shear=shear, resample=Image.NEAREST, fillcolor=None)
if (np.random.rand() < self.p):
im1 = transforms.functional.hflip(im1)
im2 = transforms.functional.hflip(im2)
gt = transforms.functional.hflip(gt)
im1 = self.normalize_transform( transforms.ToTensor()(im1).float() )
im2 = self.normalize_transform( transforms.ToTensor()(im2).float() )
# Swap randomly
if self.training and np.random.rand() < 0.5:
im2,im1 = im1,im2
gt_orig = transforms.ToTensor()(gt)
gt = prepareGTData( np.moveaxis(gt_orig.numpy()*255, 0, -1))
gt = transforms.ToTensor()(gt).long()
if self.training and np.random.rand() < 0.15:
im2 = im1
gt = torch.zeros_like(gt, dtype=torch.long)
return { 'im1': im1, 'im2':im2, 'gt': gt , 'gt_orig' : gt_orig }
def parse_args():
parser = argparse.ArgumentParser(description='Dataset Loader for PyTorch Networks')
parser.add_argument('--train_file_path', type=str, default='./train/train.txt')
return parser.parse_args()
def visualizeImages(im1,im2,gt):
mean = torch.from_numpy(np.array([0.485, 0.456, 0.406]).reshape(3,1,1)).float()
std = torch.from_numpy(np.array([0.229, 0.224, 0.225]).reshape(3,1,1)).float()
transform = transforms.ToPILImage()
c, h, w = im1.shape
combined = np.zeros((c, h, w*3))
combined[:,:, :w] = std*im1 + mean
combined[:,:, w:2*w] = std*im2 + mean
combined[:,:, 2*w:] = gt
combined = transform(torch.from_numpy(combined).float())
# im1 = transform(std*im1 + mean)
# im2 = transform(std*im2 + mean)
return combined
def visualizeAllImages(im1,im2,gt, mask):
mean = torch.from_numpy(np.array([0.485, 0.456, 0.406]).reshape(3,1,1)).float()
std = torch.from_numpy(np.array([0.229, 0.224, 0.225]).reshape(3,1,1)).float()
transform = transforms.ToPILImage()
c, h, w = im1.shape
combined = np.zeros((c, h, w*4))
combined[:,:, :w] = std*im1 + mean
combined[:,:, w:2*w] = std*im2 + mean
combined[:,:, 2*w:3*w] = gt
combined[:,:, 3*w:] = mask
combined = transform(torch.from_numpy(combined).float())
# im1 = transform(std*im1 + mean)
# im2 = transform(std*im2 + mean)
return combined
def visualizeMaskedImages(im1,im2,gt, mask):
mean = torch.from_numpy(np.array([0.485, 0.456, 0.406]).reshape(3,1,1)).float()
std = torch.from_numpy(np.array([0.229, 0.224, 0.225]).reshape(3,1,1)).float()
transform = transforms.ToPILImage()
bin_mask = mask == 1 # no-change is (1,1,1)
bin_mask = bin_mask.double()
alpha = 0.4
beta = (1.0 - alpha)
c, h, w = im1.shape
combined = np.zeros((c, h, w*4))
combined[:,:, :w] = std*im1 + mean
combined[:,:, w:2*w] = std*im2 + mean
combined[:,:, 2*w:3*w] = gt
combined[:,:, 3*w:] = combined[:,:, w:2*w]*bin_mask + (1-bin_mask)*(alpha*combined[:,:, w:2*w] + beta*(mask) )
combined = transform(torch.from_numpy(combined).float())
# im1 = transform(std*im1 + mean)
# im2 = transform(std*im2 + mean)
return combined
def visualizeOneImage(im1):
mean = torch.from_numpy(np.array([0.485, 0.456, 0.406]).reshape(3,1,1)).float()
std = torch.from_numpy(np.array([0.229, 0.224, 0.225]).reshape(3,1,1)).float()
transform = transforms.ToPILImage()
im1 = transform(std*im1 + mean)
return im1
def labelVisualize(img, num_class=10, target_size = (224,224)):
#img = (img.reshape( num_class, target_size[0], target_size[1]) ).argmax( axis=0 )
seg_img = np.zeros( ( target_size[0], target_size[1], 3 ) ).astype('uint8')
for c in range(num_class):
seg_img[:,:,0] += ( (img[:,: ] == c )*( CLASS_DICT[c][0] )).astype('uint8')
seg_img[:,:,1] += ((img[:,: ] == c )*( CLASS_DICT[c][1] )).astype('uint8')
seg_img[:,:,2] += ((img[:,: ] == c )*( CLASS_DICT[c][2] )).astype('uint8')
#seg_img = io.transform.resize(seg_img , (768, 1024 ))
#skio.imsave( "output/output" + str(imNumber) + ".png" , seg_img )
return seg_img
def main(args):
vlcmu = VLCMUCDDataset(args.train_file_path, training=True, probability=0.5)
for idx in range(len(vlcmu)):
print(idx)
data = vlcmu[idx]
# print(data['gt'].shape)
# plt.imshow(visualizeImages(data['im1'],data['im2'],data['gt_orig']))
edges = data['edges']
plt.imshow(edges)
plt.draw()
plt.pause(0.3)
if __name__ == '__main__':
args = parse_args()
main(args)