-
Notifications
You must be signed in to change notification settings - Fork 30
/
data_loader.py
246 lines (227 loc) · 11.2 KB
/
data_loader.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
from __future__ import division
import os
import random
import tensorflow as tf
class DataLoader(object):
def __init__(self,
dataset_dir=None,
batch_size=None,
img_height=None,
img_width=None,
num_source=None,
num_scales=None):
self.dataset_dir = dataset_dir
self.batch_size = batch_size
self.img_height = img_height
self.img_width = img_width
self.num_source = num_source
self.num_scales = num_scales
def load_train_batch(self, is_training=True):
"""Load a batch of training instances.
"""
seed = random.randint(0, 2**31 - 1)
# Load the list of training files into queues
file_list = self.format_file_list(self.dataset_dir, 'train')
image_paths_queue = tf.train.string_input_producer(
file_list['image_file_list'],
seed=seed,
shuffle=True)
cam_paths_queue = tf.train.string_input_producer(
file_list['cam_file_list'],
seed=seed,
shuffle=True)
self.steps_per_epoch = int(
len(file_list['image_file_list'])//self.batch_size)
# Load images
img_reader = tf.WholeFileReader()
_, image_contents = img_reader.read(image_paths_queue)
image_seq = tf.image.decode_jpeg(image_contents)
tgt_image, src_image_stack = \
self.unpack_image_sequence(
image_seq, self.img_height, self.img_width, self.num_source)
# Load camera intrinsics
cam_reader = tf.TextLineReader()
_, raw_cam_contents = cam_reader.read(cam_paths_queue)
rec_def = []
for i in range(9):
rec_def.append([1.])
raw_cam_vec = tf.decode_csv(raw_cam_contents,
record_defaults=rec_def)
raw_cam_vec = tf.stack(raw_cam_vec)
intrinsics = tf.reshape(raw_cam_vec, [3, 3])
# Form training batches
src_image_stack, tgt_image, intrinsics = \
tf.train.batch([src_image_stack, tgt_image, intrinsics],
batch_size=self.batch_size)
# Data augmentation
image_all = tf.concat([tgt_image, src_image_stack], axis=3)
image_all, intrinsics, image_augall = self.data_augmentation(
image_all, intrinsics, self.img_height, self.img_width)
tgt_image = image_all[:, :, :, :3]
src_image_stack = image_all[:, :, :, 3:]
tgt_image_aug = image_augall[:, :, :, :3]
src_image_stack_aug = image_augall[:, :, :, 3:]
intrinsics = self.get_multi_scale_intrinsics(
intrinsics, self.num_scales)
if is_training:
return tgt_image, src_image_stack, intrinsics, tgt_image_aug, src_image_stack_aug
else:
return tgt_image, src_image_stack, intrinsics
def make_intrinsics_matrix(self, fx, fy, cx, cy):
# Assumes batch input
batch_size = fx.get_shape().as_list()[0]
zeros = tf.zeros_like(fx)
r1 = tf.stack([fx, zeros, cx], axis=1)
r2 = tf.stack([zeros, fy, cy], axis=1)
r3 = tf.constant([0.,0.,1.], shape=[1, 3])
r3 = tf.tile(r3, [batch_size, 1])
intrinsics = tf.stack([r1, r2, r3], axis=1)
return intrinsics
def data_augmentation(self, im, intrinsics, out_h, out_w):
# Random scaling
def random_scaling(im, intrinsics):
batch_size, in_h, in_w, _ = im.get_shape().as_list()
scaling = tf.random_uniform([2], 1, 1.15)
x_scaling = scaling[0]
y_scaling = scaling[1]
out_h = tf.cast(in_h * y_scaling, dtype=tf.int32)
out_w = tf.cast(in_w * x_scaling, dtype=tf.int32)
im = tf.image.resize_area(im, [out_h, out_w])
fx = intrinsics[:,0,0] * x_scaling
fy = intrinsics[:,1,1] * y_scaling
cx = intrinsics[:,0,2] * x_scaling
cy = intrinsics[:,1,2] * y_scaling
intrinsics = self.make_intrinsics_matrix(fx, fy, cx, cy)
return im, intrinsics
# Random cropping
def random_cropping(im, intrinsics, out_h, out_w):
# batch_size, in_h, in_w, _ = im.get_shape().as_list()
batch_size, in_h, in_w, _ = tf.unstack(tf.shape(im))
offset_y = tf.random_uniform([1], 0, in_h - out_h + 1, dtype=tf.int32)[0]
offset_x = tf.random_uniform([1], 0, in_w - out_w + 1, dtype=tf.int32)[0]
im = tf.image.crop_to_bounding_box(
im, offset_y, offset_x, out_h, out_w)
fx = intrinsics[:,0,0]
fy = intrinsics[:,1,1]
cx = intrinsics[:,0,2] - tf.cast(offset_x, dtype=tf.float32)
cy = intrinsics[:,1,2] - tf.cast(offset_y, dtype=tf.float32)
intrinsics = self.make_intrinsics_matrix(fx, fy, cx, cy)
return im, intrinsics
# Random photometric augmentation
# Credit: https://github.com/simonmeister/UnFlow/blob/master/src/e2eflow/core/augment.py
def random_photometric(im, noise_stddev=0.04, min_contrast=-0.2, max_contrast=0.2, brightness_stddev=0.02, min_colour=0.9, max_colour=1.1, min_gamma=0.8, max_gamma=1.2):
"""
Applies photometric augmentations to a list of image batches.
Args:
im: list of 3-channel image batches normalized to [0, 1].
Returns:
Batch of normalized images with photometric augmentations. Has the same shape as the input batch.
"""
batch_size, in_h, in_w, _ = im[0].get_shape().as_list()
contrast = tf.random_uniform([batch_size, 1], min_contrast, max_contrast)
gamma = tf.random_uniform([batch_size, 1], min_gamma, max_gamma)
gamma_inv = 1.0/gamma
colour = tf.random_uniform([batch_size, 3], min_colour, max_colour)
if noise_stddev > 0.0:
noise = tf.random_normal([batch_size, 1], stddev=noise_stddev)
else:
noise = tf.zeros([batch_size, 1])
if brightness_stddev > 0.0:
brightness = tf.random_normal([batch_size, 1], stddev=brightness_stddev)
else:
brightness = tf.zeros([batch_size, 1])
out = []
for temp in im:
# Transpose to [height, width, num_batch, channels]
im_re = tf.transpose(temp, [1, 2, 0, 3])
im_re = (im_re * (contrast + 1.0) + brightness) * colour
im_re = tf.maximum(0.0, tf.minimum(1.0, im_re))
im_re = tf.pow(im_re, gamma_inv)
im_re = im_re + noise
im_re = tf.maximum(0.0, tf.minimum(1.0, im_re))
temp = tf.transpose(im_re, [2, 0, 1, 3])
temp = tf.stop_gradient(temp)
out.append(temp)
return tf.concat(out, axis=-1)
im, intrinsics = random_scaling(im, intrinsics)
im, intrinsics = random_cropping(im, intrinsics, out_h, out_w)
# [0, 255] -> [0, 1]
im_photo = im/255.
im_photo = [im_photo[:,:,:,3*i:3*(i+1)] for i in range(self.num_source+1)]
im_photo = random_photometric(im_photo)
# [0, 1] -> [0, 255]
im_photo = im_photo*255.
im = tf.cast(im, dtype=tf.uint8)
im_photo = tf.cast(im_photo, dtype=tf.uint8)
return im, intrinsics, im_photo
def format_file_list(self, data_root, split):
with open(data_root + '/%s.txt' % split, 'r') as f:
frames = f.readlines()
subfolders = [x.split(' ')[0] for x in frames]
frame_ids = [x.split(' ')[1][:-1] for x in frames]
image_file_list = [os.path.join(data_root, subfolders[i],
frame_ids[i] + '.jpg') for i in range(len(frames))]
cam_file_list = [os.path.join(data_root, subfolders[i],
frame_ids[i] + '_cam.txt') for i in range(len(frames))]
all_list = {}
all_list['image_file_list'] = image_file_list
all_list['cam_file_list'] = cam_file_list
return all_list
def unpack_image_sequence(self, image_seq, img_height, img_width, num_source):
# Assuming the center image is the target frame
tgt_start_idx = int(img_width * (num_source//2))
tgt_image = tf.slice(image_seq,
[0, tgt_start_idx, 0],
[-1, img_width, -1])
# Source frames before the target frame
src_image_1 = tf.slice(image_seq,
[0, 0, 0],
[-1, int(img_width * (num_source//2)), -1])
# Source frames after the target frame
src_image_2 = tf.slice(image_seq,
[0, int(tgt_start_idx + img_width), 0],
[-1, int(img_width * (num_source//2)), -1])
src_image_seq = tf.concat([src_image_1, src_image_2], axis=1)
# Stack source frames along the color channels (i.e. [H, W, N*3])
src_image_stack = tf.concat([tf.slice(src_image_seq,
[0, i*img_width, 0],
[-1, img_width, -1])
for i in range(num_source)], axis=2)
src_image_stack.set_shape([img_height,
img_width,
num_source * 3])
tgt_image.set_shape([img_height, img_width, 3])
return tgt_image, src_image_stack
def batch_unpack_image_sequence(self, image_seq, img_height, img_width, num_source):
# Assuming the center image is the target frame
tgt_start_idx = int(img_width * (num_source//2))
tgt_image = tf.slice(image_seq,
[0, 0, tgt_start_idx, 0],
[-1, -1, img_width, -1])
# Source frames before the target frame
src_image_1 = tf.slice(image_seq,
[0, 0, 0, 0],
[-1, -1, int(img_width * (num_source//2)), -1])
# Source frames after the target frame
src_image_2 = tf.slice(image_seq,
[0, 0, int(tgt_start_idx + img_width), 0],
[-1, -1, int(img_width * (num_source//2)), -1])
src_image_seq = tf.concat([src_image_1, src_image_2], axis=2)
# Stack source frames along the color channels (i.e. [B, H, W, N*3])
src_image_stack = tf.concat([tf.slice(src_image_seq,
[0, 0, i*img_width, 0],
[-1, -1, img_width, -1])
for i in range(num_source)], axis=3)
return tgt_image, src_image_stack
def get_multi_scale_intrinsics(self, intrinsics, num_scales):
intrinsics_mscale = []
# Scale the intrinsics accordingly for each scale
for s in range(num_scales):
fx = intrinsics[:,0,0]/(2 ** s)
fy = intrinsics[:,1,1]/(2 ** s)
cx = intrinsics[:,0,2]/(2 ** s)
cy = intrinsics[:,1,2]/(2 ** s)
intrinsics_mscale.append(
self.make_intrinsics_matrix(fx, fy, cx, cy))
intrinsics_mscale = tf.stack(intrinsics_mscale, axis=1)
return intrinsics_mscale