-
Notifications
You must be signed in to change notification settings - Fork 7
/
modules.py
339 lines (284 loc) · 17 KB
/
modules.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
'''
Building blocks implemented in Tensorflow
-Cross-modal flow estimation
|-FeatureExtractor
|-FlowEstimation
|-pose_module
|-linear_pose_constraint
|-pose_to_flow
|-PoseFusion
-RGB-D module depth refinement
|-RGBToFKPN
Di Qiu, 13-06-2019
'''
import numpy as np
import tensorflow as tf
from utils.utils import LeakyReLU, pad, antipad
from keras.layers.core import Dense, Flatten, Lambda
from keras.layers.normalization import BatchNormalization
from utils.correlation import correlation
from utils.camera_util import *
from utils.flowlib import *
slim = tf.contrib.slim
##### Start of cross-modal flow estimation #####
class FeatureExtractor(object):
'''
Feature extraction tower for FlowNetC
FlowNetC implementation taken from: https://github.com/sampepose/flownet2-tf
'''
def __init__(self, reuse=False):
self.reuse = reuse
def __call__(self, input):
with slim.arg_scope([slim.conv2d, slim.conv2d_transpose],
# Only backprop this network if trainable
trainable=True,
# He (aka MSRA) weight initialization
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=LeakyReLU,
# We will do our own padding to match the original Caffe code
padding='VALID'):
weights_regularizer = slim.l2_regularizer(0.0004)
with slim.arg_scope([slim.conv2d], weights_regularizer=weights_regularizer):
with slim.arg_scope([slim.conv2d], stride=2):
conv_a_1 = slim.conv2d(pad(input, 3), 64, 7, scope='FlowNetC/conv1', reuse=self.reuse)
conv_a_2 = slim.conv2d(pad(conv_a_1, 2), 128, 5, scope='FlowNetC/conv2', reuse=self.reuse)
conv_a_3 = slim.conv2d(pad(conv_a_2, 2), 256, 5, scope='FlowNetC/conv3', reuse=self.reuse)
return {'conv_1': conv_a_1,
'conv_2': conv_a_2,
'conv_3': conv_a_3
}
class FlowEstimator(object):
'''
FlowNetC implementation taken from: https://github.com/sampepose/flownet2-tf
'''
def __init__(self, height, width, reuse=False):
self.reuse = reuse
self.height = height
self.width = width
def __call__(self, featureA, featureB):
conv_a_2 = featureA['conv_2']
conv_a_3 = featureA['conv_3']
conv_b_3 = featureB['conv_3']
# print(conv_a_3.get_shape().as_list())
# print(conv_b_3.get_shape().as_list())
with slim.arg_scope([slim.conv2d, slim.conv2d_transpose],
# Only backprop this network if trainable
trainable=True,
# He (aka MSRA) weight initialization
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=LeakyReLU,
# We will do our own padding to match the original Caffe code
padding='VALID'):
cc = correlation(conv_a_3, conv_b_3, 1, 20, 1, 2, 20)
cc_relu = LeakyReLU(cc)
# Combine cross correlation results with convolution of feature map A
net_conv = slim.conv2d(conv_a_3, 32, 1, scope='FlowNetC/conv_redir', reuse=self.reuse)
# Concatenate along the channels axis
net = tf.concat([net_conv, cc_relu], axis=3)
conv3_1 = slim.conv2d(pad(net), 256, 3, scope='FlowNetC/conv3_1', reuse=self.reuse)
# print(conv3_1.get_shape().as_list())
with slim.arg_scope([slim.conv2d], num_outputs=512, kernel_size=3, reuse=self.reuse):
conv4 = slim.conv2d(pad(conv3_1), stride=2, scope='FlowNetC/conv4')
conv4_1 = slim.conv2d(pad(conv4), scope='FlowNetC/conv4_1')
conv5 = slim.conv2d(pad(conv4_1), stride=2, scope='FlowNetC/conv5')
conv5_1 = slim.conv2d(pad(conv5), scope='FlowNetC/conv5_1')
conv6 = slim.conv2d(pad(conv5_1), 1024, 3, stride=2, scope='FlowNetC/conv6', reuse=self.reuse)
# print(conv6.get_shape().as_list())
conv6_1 = slim.conv2d(pad(conv6), 1024, 3, scope='FlowNetC/conv6_1', reuse=self.reuse)
""" START: Refinement Network """
with slim.arg_scope([slim.conv2d], reuse=self.reuse):
with slim.arg_scope([slim.conv2d_transpose], biases_initializer=None, reuse=self.reuse):
predict_flow6 = slim.conv2d(pad(conv6_1), 2, 3,
scope='FlowNetC/predict_flow6',
activation_fn=None)
deconv5 = antipad(slim.conv2d_transpose(conv6, 512, 4,
stride=2,
scope='FlowNetC/deconv5'))
upsample_flow6to5 = antipad(slim.conv2d_transpose(predict_flow6, 2, 4,
stride=2,
scope='FlowNetC/upsample_flow6to5',
activation_fn=None))
concat5 = tf.concat([conv5_1, deconv5, upsample_flow6to5], axis=3)
predict_flow5 = slim.conv2d(pad(concat5), 2, 3,
scope='FlowNetC/predict_flow5',
activation_fn=None)
deconv4 = antipad(slim.conv2d_transpose(concat5, 256, 4,
stride=2,
scope='FlowNetC/deconv4'))
upsample_flow5to4 = antipad(slim.conv2d_transpose(predict_flow5, 2, 4,
stride=2,
scope='FlowNetC/upsample_flow5to4',
activation_fn=None))
concat4 = tf.concat([conv4_1, deconv4, upsample_flow5to4], axis=3)
predict_flow4 = slim.conv2d(pad(concat4), 2, 3,
scope='FlowNetC/predict_flow4',
activation_fn=None)
deconv3 = antipad(slim.conv2d_transpose(concat4, 128, 4,
stride=2,
scope='FlowNetC/deconv3'))
upsample_flow4to3 = antipad(slim.conv2d_transpose(predict_flow4, 2, 4,
stride=2,
scope='FlowNetC/upsample_flow4to3',
activation_fn=None))
concat3 = tf.concat([conv3_1, deconv3, upsample_flow4to3], axis=3)
predict_flow3 = slim.conv2d(pad(concat3), 2, 3,
scope='FlowNetC/predict_flow3',
activation_fn=None)
deconv2 = antipad(slim.conv2d_transpose(concat3, 64, 4,
stride=2,
scope='FlowNetC/deconv2'))
upsample_flow3to2 = antipad(slim.conv2d_transpose(predict_flow3, 2, 4,
stride=2,
scope='FlowNetC/upsample_flow3to2',
activation_fn=None))
concat2 = tf.concat([conv_a_2, deconv2, upsample_flow3to2], axis=3)
predict_flow2 = slim.conv2d(pad(concat2), 2, 3,
scope='FlowNetC/predict_flow2',
activation_fn=None)
""" END: Refinement Network """
flow = predict_flow2
# TODO: Look at Accum (train) or Resample (deploy) to see if we need to do something different
flow = tf.image.resize_bilinear(flow,
tf.stack([self.height, self.width]),
align_corners=True)
return {
'predict_flows': [predict_flow6, predict_flow5, predict_flow4, predict_flow3, predict_flow2],
'flow': flow,
'feature': conv3_1
}
def pose_module(D, rf_flow, conf_):
'''
:param D:
:param rf_predictions:
:param conf_:
:return:
'''
b_, h_, w_, _ = D.get_shape().as_list()
pose = linear_pose_constraint(D, rf_flow, conf_)
pose_flow = pose_to_flow(pose, D)
return pose_flow, pose
def linear_pose_constraint(D, flow, weight):
# build left hand matrix
nD = D/1000. # some scale change
nD = tf.squeeze(nD)
b, h, w, _ = flow.get_shape().as_list()
D_vec = tf.reshape(nD, [b, h*w, 1])
weight = tf.reshape(weight, [b, h*w, 1])
M = tf.concat([D_vec, tf.ones([b, h*w, 1], 'float32')], 2) * weight
# build right hand side
dflowx = nD * flow[:, :, :, 0]
dflowy = nD * flow[:, :, :, 1]
dflowx = tf.reshape(dflowx, [b, h*w, 1]) * weight
dflowy = tf.reshape(dflowy, [b, h*w, 1]) * weight
rhs = tf.concat([dflowx, dflowy], 2)
sol = tf.matrix_solve_ls(M/10000., rhs/10000., l2_regularizer=0.00001) # some other stupid redundant scale change...
return {'c': sol[:, 0, :], 't': sol[:, 1, :] * 1000.}
def pose_to_flow(pose, D):
b_, h_, w_, _ = D.get_shape().as_list()
c = pose['c']
c = tf.reshape(c, [b_, 1, 1, 2])
t = pose['t']
t = tf.reshape(t, [b_, 1, 1, 2])
pose_flow = c + t / D
return pose_flow
class PoseFusion(object):
def __init__(self, reuse=False):
self.reuse = reuse
def __call__(self, pose_flow, flow):
with slim.arg_scope([slim.conv2d, slim.conv2d_transpose],
# Only backprop this network if trainable
trainable=True,
# He (aka MSRA) weight initialization
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=LeakyReLU,
weights_regularizer=slim.l2_regularizer(0.0004)):
inputs = tf.concat([pose_flow, flow], 3)
conv0 = slim.conv2d(inputs, 64, 3, stride=1, scope='pose/conv0', padding='SAME')
conv1 = slim.conv2d(conv0, 64, 3, stride=2, scope='pose/conv1', padding='SAME')
conv1_1 = slim.conv2d(conv1, 128, 3, stride=1, scope='pose/conv1_1', padding='SAME')
conv2 = slim.conv2d(conv1_1, 128, 3, stride=2, scope='pose/conv2', padding='SAME')
conv2_1 = slim.conv2d(conv2, 128, 3, stride=1, scope='pose/conv2_1', padding='SAME')
predict_flow2 = slim.conv2d(conv2_1, 2, 3, stride=1, scope='pose/predict_flow2', padding='SAME')
upconv1 = slim.conv2d_transpose(tf.concat([predict_flow2, conv2_1], 3), 128, 4, stride=2, scope='pose/upconv1', padding='SAME')
_, h1_, w1_, _ = upconv1.get_shape().as_list()
rconv1 = slim.conv2d(tf.concat([upconv1, conv1_1], 3), 64, 3,
stride=1, scope='pose/rconv1', padding='SAME')
predict_flow1 = slim.conv2d(rconv1, 2, 3, stride=1, scope='pose/predict_flow1', padding='SAME')
upconv0 = slim.conv2d_transpose(tf.concat([predict_flow1, rconv1] ,3), 64, 4, stride=2, scope='pose/upconv0', padding='SAME')
_, h0_, w0_, _ = upconv0.get_shape().as_list()
rconv0 = slim.conv2d(tf.concat([upconv0, conv0], 3), 64, 3,
stride=1, scope='pose/rconv0', padding='SAME')
predict_flow0 = slim.conv2d(rconv0, 2, 3, stride=1, scope='pose/predict_flow0', padding='SAME')
return [predict_flow2, predict_flow1, predict_flow0]
##### End of cross-modal flow estimation #####
##### Start of RGB-D depth refinement #####
class RGBToFKPN(object):
def __init__(self, ks=3):
self.ks = ks
def __call__(self, flowD, R, flowL):
nch = 64
nrep = 1
ks = self.ks
R = R - tf.reduce_mean(R, [1, 2, 3], keep_dims=True)
flowL = tf.tile(flowL, [1, 1, 1, 3])
inputs = tf.concat([flowD, flowL, R], 3)
#inputs = tf.concat([flowD, flowL], 3)
_, height, width, _ = flowD.get_shape().as_list()
with slim.arg_scope([slim.conv2d, slim.conv2d_transpose],
# Only backprop this network if trainable
trainable=True,
# He (aka MSRA) weight initialization
weights_initializer=slim.variance_scaling_initializer(), # biases_initializer=None,
# normalizer_fn=slim.batch_norm,
# tf.truncated_normal_initializer(stddev=0.1), #slim.variance_scaling_initializer(),
# tf.contrib.layers.xavier_initializer(uniform=True, dtype=tf.float32),
activation_fn=LeakyReLU,
padding='SAME'):
# weights_regularizer=slim.l2_regularizer(0.00004)):
with tf.variable_scope('dynamic'):
dconv0 = slim.conv2d(inputs, nch, 3, stride=1, scope='ddk/conv0', padding='SAME')
dconv1 = slim.repeat(dconv0, nrep, slim.conv2d, nch, [3, 3], stride=1, scope='ddk/conv1', padding='SAME')
# print(dconv1.get_shape().as_list())
dconv2 = slim.conv2d(dconv1, nch * 2, 3, stride=2, scope='ddk/conv2', padding='SAME')
dconv2_1 = slim.repeat(dconv2, nrep, slim.conv2d, nch * 2, [3, 3], stride=1, scope='ddk/conv2_1',
padding='SAME')
# print(dconv2_1.get_shape().as_list())
dconv3 = slim.conv2d(dconv2_1, nch * 2, 3, stride=2, scope='ddk/conv3', padding='SAME')
dconv3_1 = slim.repeat(dconv3, nrep, slim.conv2d, nch * 2, [3, 3], stride=1, scope='ddk/conv3_1',
padding='SAME')
# print(dconv3_1.get_shape().as_list())
dconv4 = slim.conv2d(dconv3_1, nch * 4, 3, stride=2, scope='ddk/conv4', padding='SAME')
dconv4_1 = slim.repeat(dconv4, nrep, slim.conv2d, nch * 4, [3, 3], stride=1, scope='ddk/conv4_1',
padding='SAME')
# print(dconv4_1.get_shape().as_list())
# UP 1
ddeconv_n1 = slim.conv2d_transpose(tf.concat([dconv4_1], 3), nch * 2, 3, stride=2,
scope='ddk/deconv_n1', padding='SAME')
ddeconv_n1_1 = slim.repeat(tf.concat([ddeconv_n1], 3), nrep, slim.conv2d, nch * 2, [3, 3], stride=1,
scope='ddk/deconv_n1_1', padding='SAME')
# print(ddeconv_n1_1.get_shape().as_list())
#
# UP 2
ddeconv_n2 = slim.conv2d_transpose(tf.concat([ddeconv_n1_1, dconv3_1], 3), nch * 2, 3, stride=2,
scope='ddk/deconv_n2', padding='SAME')
ddeconv_n2_1 = slim.repeat(tf.concat([ddeconv_n2], 3), nrep, slim.conv2d, nch * 2, [3, 3], stride=1,
scope='ddk/deconv_n2_1', padding='SAME')
# print(ddeconv_n2_1.get_shape().as_list())
# pwd2 = slim.conv2d_transpose(ddeconv_n2_1, nch, 3, stride=2,
# UP 3
ddeconv_n3 = slim.conv2d_transpose(tf.concat([ddeconv_n2_1, dconv2_1], 3), nch, 3, stride=2,
scope='ddk/deconv_n3', padding='SAME')
ddeconv_n3_1 = slim.repeat(tf.concat([dconv1, ddeconv_n3], 3), nrep, slim.conv2d, nch, [3, 3], stride=1,
scope='ddk/deconv_n3_1', padding='SAME')
# print(ddeconv_n3_1.get_shape().as_list())
# Apply kpn
wb = slim.repeat(ddeconv_n3_1, nrep, slim.conv2d, ks ** 2 + 1, [3, 3], scope='ddk/w', padding='SAME',
biases_initializer=None, activation_fn=None)
w = wb[:, :, :, :-1]
b = tf.expand_dims(wb[:, :, :, -1], 3)
w = w / tf.reduce_sum(tf.abs(w) + 1e-6, 3, keep_dims=True)
img_patch = tf.extract_image_patches(flowD+b , [1, ks, ks, 1], [1, 1, 1, 1], [1, 1, 1, 1], "SAME")
filtimg = tf.reduce_sum((img_patch) * w, 3, keep_dims=True)
# filtimg = tf.image.resize_bilinear(pfiltimg3, [H, W])
return [filtimg, w, b] # [dx, dy, w, b]
##### End of RGB-D module depth refinement #####