/
dataset.py
191 lines (157 loc) · 6.99 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
import numpy as np
import io
from PIL import Image
import torch
import torchvision.transforms as transforms
from torch.utils.data import Dataset
from utils.flowlib import read_flo_file
from utils import image_crop, image_resize, image_flow_crop, image_flow_resize, flow_sampler, image_flow_aug, flow_aug
class ColorAugmentation(object):
def __init__(self, eig_vec=None, eig_val=None):
if eig_vec == None:
eig_vec = torch.Tensor([
[ 0.4009, 0.7192, -0.5675],
[-0.8140, -0.0045, -0.5808],
[ 0.4203, -0.6948, -0.5836],
])
if eig_val == None:
eig_val = torch.Tensor([[0.2175, 0.0188, 0.0045]])
self.eig_val = eig_val # 1*3
self.eig_vec = eig_vec # 3*3
def __call__(self, tensor):
assert tensor.size(0) == 3
alpha = torch.normal(means=torch.zeros_like(self.eig_val))*0.1
quatity = torch.mm(self.eig_val*alpha, self.eig_vec)
tensor = tensor + quatity.view(3, 1, 1)
return tensor
def pil_loader(img_str, ch):
buff = io.BytesIO(img_str)
if ch == 1:
return Image.open(buff)
else:
with Image.open(buff) as img:
img = img.convert('RGB')
return img
def pil_loader_str(img_str, ch):
if ch == 1:
return Image.open(img_str)
else:
with Image.open(img_str) as img:
img = img.convert('RGB')
return img
class ImageFlowDataset(Dataset):
def __init__(self, meta_file, config, phase):
self.img_transform = transforms.Compose([
transforms.Normalize(config['data_mean'], config['data_div'])
])
print("building dataset from {}".format(meta_file))
self.flow_file_type = config['flow_file_type']
self.metas = []
self.num = 0
for mf in meta_file:
with open(mf, 'r') as f:
lines = f.readlines()
self.num += len(lines)
for line in lines:
if self.flow_file_type == "flo":
img0_path, img1_path, flow_path = line.rstrip().split()
self.metas.append((img0_path, img1_path, flow_path))
elif self.flow_file_type == "jpg":
img0_path, img1_path, flow_path_x, flow_path_y = line.rstrip().split()
self.metas.append((img0_path, img1_path, flow_path_x, flow_path_y))
else:
raise Exception("No such flow_file_type: {}".format(self.flow_file_type))
print("read meta done, total: {}".format(self.num))
self.phase = phase
self.short_size = config.get('short_size', None)
self.long_size = config.get('long_size', None)
self.crop_size = config.get('crop_size', None)
self.sample_strategy = config['sample_strategy']
self.sample_bg_ratio = config['sample_bg_ratio']
self.nms_ks = config['nms_ks']
self.max_num_guide = config['max_num_guide']
if self.phase == "train":
self.aug_flip = config['image_flow_aug'].get('flip', False)
self.aug_reverse = config['flow_aug'].get('reverse', False)
self.aug_scale = config['flow_aug'].get('scale', False)
self.aug_rotate = config['flow_aug'].get('rotate', False)
def __len__(self):
return self.num
def __getitem__(self, idx):
img1_fn = self.metas[idx][0]
img2_fn = self.metas[idx][1]
if self.flow_file_type == 'flo':
flowname = self.metas[idx][2]
flow = read_flo_file(flowname) # h, w, 2
else:
flownamex = self.metas[idx][2]
flownamey = self.metas[idx][3]
flowx = np.array(Image.open(flownamex)).astype(np.float32) / 255 * 100 - 50
flowy = np.array(Image.open(flownamey)).astype(np.float32) / 255 * 100 - 50
flow = np.concatenate((flowx[:,:,np.newaxis], flowy[:,:,np.newaxis]), axis=2)
img1 = pil_loader_str(img1_fn, ch=3)
img2 = pil_loader_str(img2_fn, ch=3)
## check size
assert img1.height == flow.shape[0]
assert img1.width == flow.shape[1]
assert img2.height == flow.shape[0]
assert img2.width == flow.shape[1]
## resize
if self.short_size is not None or self.long_size is not None:
img1, img2, flow, ratio = image_flow_resize(
img1, img2, flow, short_size=self.short_size,
long_size=self.long_size)
## crop
if self.crop_size is not None:
img1, img2, flow, offset = image_flow_crop(
img1, img2, flow, self.crop_size, self.phase)
## augmentation
if self.phase == 'train':
# image flow aug
img1, img2, flow = image_flow_aug(img1, img2, flow, flip_horizon=self.aug_flip)
# flow aug
flow = flow_aug(flow, reverse=self.aug_reverse,
scale=self.aug_scale, rotate=self.aug_rotate)
## transform
img1 = torch.from_numpy(np.array(img1).astype(np.float32).transpose((2,0,1)))
img2 = torch.from_numpy(np.array(img2).astype(np.float32).transpose((2,0,1)))
img1 = self.img_transform(img1)
img2 = self.img_transform(img2)
## sparse sampling
sparse_flow, mask = flow_sampler(
flow, strategy=self.sample_strategy,
bg_ratio=self.sample_bg_ratio, nms_ks=self.nms_ks,
max_num_guide=self.max_num_guide) # (h,w,2), (h,w,2)
flow = torch.from_numpy(flow.transpose((2, 0, 1)))
sparse_flow = torch.from_numpy(sparse_flow.transpose((2, 0, 1)))
mask = torch.from_numpy(mask.transpose((2, 0, 1)).astype(np.float32))
return img1, sparse_flow, mask, flow, img2
class ImageDataset(Dataset):
def __init__(self, meta_file, config):
self.img_transform = transforms.Compose([
transforms.Normalize(config['data_mean'], config['data_div'])
])
print("building dataset from {}".format(meta_file))
with open(meta_file, 'r') as f:
lines = f.readlines()
self.num = len(lines)
self.metas = [l.rstrip() for l in lines]
print("read meta done, total: {}".format(self.num))
self.short_size = config.get('short_size', None)
self.long_size = config.get('long_size', None)
self.crop_size = config.get('crop_size', None)
def __len__(self):
return self.num
def __getitem__(self, idx):
img_fn = self.metas[idx]
img = pil_loader_str(img_fn, ch=3)
## resize
if self.short_size is not None or self.long_size is not None:
img, size = image_resize(img, short_size=self.short_size, long_size=self.long_size)
## crop
if self.crop_size is not None:
img, offset = image_crop(img, self.crop_size)
## transform
img = torch.from_numpy(np.array(img).astype(np.float32).transpose((2,0,1)))
img = self.img_transform(img)
return img, torch.LongTensor([idx]), torch.LongTensor(offset), torch.LongTensor(size)