-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathmodules.py
187 lines (147 loc) · 10.7 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
import tensorflow as tf
class CostVolumeLayer(tf.Module):
def __init__(self, search_range=4):
super(CostVolumeLayer, self).__init__(name='cost_volume')
self.window = search_range
def __call__(self, x, warped, dimension=2, cam_signals=None):
assert dimension in [1, 2]
total = []
keys = []
if dimension == 1:
row_shifted = warped
for i in range(2 * self.window + 1):
if i != 0:
if cam_signals is not None and not isinstance(cam_signals,int):
outputs = []
for s_index in range(len(row_shifted)):
output = row_shifted[s_index][tf.newaxis]
cam_signal = cam_signals[s_index]
if int(cam_signal) == -1:
output = tf.pad(
output, [[0, 0], [0, 0], [0, 1], [0, 0]])
output = tf.keras.layers.Cropping2D([[0, 0], [1, 0]])(
output) # [[top_crop, bottom_crop], [left_crop, right_crop]]
else:
output = tf.pad(
output, [[0, 0],[0, 0], [1, 0], [0, 0]])
output = tf.keras.layers.Cropping2D([[0, 0], [0, 1]])(
output) # [[top_crop, bottom_crop], [left_crop, right_crop]]
outputs.append(output)
row_shifted = tf.squeeze(tf.stack(outputs, axis=1), axis=0)
else:
row_shifted = tf.pad(row_shifted, [[0, 0], [0, 0], [1, 0], [0, 0]])
row_shifted = tf.keras.layers.Cropping2D([[0, 0], [0, 1]])(row_shifted)
total.append(tf.reduce_mean(row_shifted * x, axis=-1))
stacked = tf.stack(total, axis=3)
return stacked / (2 * self.window + 1)
else: # dimension == 2
row_shifted = [warped]
for i in range(self.window+1):
if i != 0:
row_shifted = [tf.pad(row_shifted[0], [[0, 0], [0, 1], [0, 0], [0, 0]]),
tf.pad(row_shifted[1], [[0, 0], [1, 0], [0, 0], [0, 0]])]
row_shifted = [tf.keras.layers.Cropping2D([[1, 0], [0, 0]])(row_shifted[0]),
tf.keras.layers.Cropping2D([[0, 1], [0, 0]])(row_shifted[1])]
for side in range(len(row_shifted)):
total.append(tf.reduce_mean(
row_shifted[side] * x, axis=-1))
keys.append([i * (-1) ** side, 0])
col_previous = [row_shifted[side], row_shifted[side]]
for j in range(1, self.window+1):
col_shifted = [tf.pad(col_previous[0], [[0, 0], [0, 0], [0, 1], [0, 0]]),
tf.pad(col_previous[1], [[0, 0], [0, 0], [1, 0], [0, 0]])]
col_shifted = [tf.keras.layers.Cropping2D([[0, 0], [1, 0]])(col_shifted[0]),
tf.keras.layers.Cropping2D([[0, 0], [0, 1]])(col_shifted[1])]
for col_side in range(len(col_shifted)):
total.append(tf.reduce_mean(
col_shifted[col_side] * x, axis=-1))
keys.append(
[i * (-1) ** side, j * (-1) ** col_side])
col_previous = col_shifted
if i == 0:
row_shifted *= 2
total = [t for t, _ in sorted(
zip(total, keys), key=lambda pair: pair[1])]
stacked = tf.stack(total, axis=3)
return stacked / ((2.0*self.window+1)**2.0)
def FeaturePyramidNetwork(image, activation_fn=tf.keras.layers.LeakyReLU(0.1), reg_constant=0):
def Conv(filters, stride, name_suffix):
return tf.keras.layers.Conv2D(filters=filters, kernel_size=3, strides=(stride, stride), padding='same', activation=activation_fn, kernel_initializer=tf.keras.initializers.Orthogonal, kernel_regularizer=tf.keras.regularizers.l2(reg_constant), name='conv'+name_suffix)
def UpConv(filters, name_suffix):
return tf.keras.layers.Conv2DTranspose(filters=filters, kernel_size=4, strides=(2, 2), padding='same', activation=activation_fn, kernel_initializer=tf.keras.initializers.Orthogonal, kernel_regularizer=tf.keras.regularizers.l2(reg_constant), name='deconv'+name_suffix)
conv1a = Conv(filters=16, stride=2, name_suffix='1a')(image)
conv1b = Conv(filters=16, stride=1, name_suffix='1b')(conv1a)
conv2a = Conv(filters=32, stride=2, name_suffix='2a')(conv1b)
conv2b = Conv(filters=32, stride=1, name_suffix='2b')(conv2a)
conv3a = Conv(filters=64, stride=2, name_suffix='3a')(conv2b)
conv3b = Conv(filters=64, stride=1, name_suffix='3b')(conv3a)
conv4a = Conv(filters=96, stride=2, name_suffix='4a')(conv3b)
conv4b = Conv(filters=96, stride=1, name_suffix='4b')(conv4a)
conv5a = Conv(filters=128, stride=2, name_suffix='5a')(conv4b)
conv5b = Conv(filters=128, stride=1, name_suffix='5b')(conv5a)
conv6a = Conv(filters=196, stride=2, name_suffix='6a')(conv5b)
conv6b = Conv(filters=196, stride=1, name_suffix='6b')(conv6a)
pyr_top = tf.keras.layers.Conv2D(filters=196, kernel_size=1, strides=(1, 1), padding='same', activation=activation_fn,
kernel_initializer=tf.keras.initializers.Orthogonal, kernel_regularizer=tf.keras.regularizers.l2(reg_constant), name='conv_pyr_top')(conv6b)
pyramid = [pyr_top]
for i, skip_features in zip(range(4, 0, -1), [conv5b, conv4b, conv3b, conv2b]):
channels = skip_features.shape.as_list()[-1]
upsampled = UpConv(filters=channels, name_suffix='_upsample' +
str(i + 1) + 'to' + str(i))(pyramid[-1])
merged_features = upsampled + skip_features
refine = Conv(filters=channels, stride=1,
name_suffix='_refine' + str(i))(merged_features)
pyramid.append(refine)
return tf.keras.Model(inputs=image, outputs=pyramid, name='feature_pyramid_network')
def SceneFlowEstimator(feature_volume, level, activation_fn=tf.keras.layers.LeakyReLU(0.1), reg_constant=0, highest_resolution=False):
def Conv(filters, name_suffix, activation=activation_fn):
return tf.keras.layers.Conv2D(filters=filters, kernel_size=3, strides=(1, 1), padding='same', activation=activation, kernel_initializer=tf.keras.initializers.Orthogonal, kernel_regularizer=tf.keras.regularizers.l2(reg_constant), name='conv'+name_suffix)
def UpConv(filters, name_suffix, activation=activation_fn):
return tf.keras.layers.Conv2DTranspose(filters=filters, kernel_size=4, strides=(2, 2), padding='same', activation=activation, kernel_initializer=tf.keras.initializers.Orthogonal, kernel_regularizer=tf.keras.regularizers.l2(reg_constant), name='deconv'+name_suffix)
conv1 = Conv(filters=128, name_suffix='1')(feature_volume)
conv2 = Conv(filters=128, name_suffix='2')(conv1)
conv3 = Conv(filters=96, name_suffix='3')(conv2)
conv4 = Conv(filters=64, name_suffix='4')(conv3)
f_lev = Conv(filters=32, name_suffix='_f')(conv4)
w_lev = Conv(filters=4, name_suffix='_w',
activation=tf.keras.activations.linear)(f_lev)
if highest_resolution:
return tf.keras.Model(inputs=feature_volume, outputs=(f_lev, w_lev), name='scene_flow_estimator_'+str(level))
else:
flow_up = UpConv(filters=4, name_suffix='_up_flow',
activation=tf.keras.activations.linear)(w_lev)
feature_up = UpConv(filters=4, name_suffix='_up_feature',
activation=tf.keras.activations.linear)(f_lev)
return tf.keras.Model(inputs=feature_volume, outputs=(w_lev, flow_up, feature_up), name='scene_flow_estimator_'+str(level))
def ContextNetwork(feature_volume, activation_fn=tf.keras.layers.LeakyReLU(0.1), reg_constant=0):
def Conv(filters, name_suffix, dilation, activation=activation_fn):
return tf.keras.layers.Conv2D(filters=filters, kernel_size=3, strides=(1, 1), padding='same', activation=activation, dilation_rate=(dilation, dilation), kernel_initializer=tf.keras.initializers.Orthogonal, kernel_regularizer=tf.keras.regularizers.l2(reg_constant), name='conv'+name_suffix)
conv1 = Conv(filters=128, name_suffix='1', dilation=1)(feature_volume)
conv2 = Conv(filters=128, name_suffix='2', dilation=2)(conv1)
conv3 = Conv(filters=128, name_suffix='3', dilation=4)(conv2)
conv4 = Conv(filters=96, name_suffix='4', dilation=8)(conv3)
conv5 = Conv(filters=64, name_suffix='5', dilation=16)(conv4)
conv6 = Conv(filters=32, name_suffix='6', dilation=1)(conv5)
conv7 = Conv(filters=4, name_suffix='7', dilation=1,
activation=tf.keras.activations.linear)(conv6)
return tf.keras.Model(inputs=feature_volume, outputs=conv7, name='context_network')
def OcclusionEstimator(features, level, activation_fn=tf.keras.layers.LeakyReLU(0.1), reg_constant=0, highest_resolution=False):
def Conv(filters, name_suffix, activation=activation_fn):
return tf.keras.layers.Conv2D(filters=filters, kernel_size=3, padding='same', activation=activation, kernel_initializer=tf.keras.initializers.Orthogonal, kernel_regularizer=tf.keras.regularizers.l2(reg_constant), name='conv'+name_suffix)
def UpConv(filters, name_suffix, activation=activation_fn):
return tf.keras.layers.Conv2DTranspose(filters=filters, kernel_size=4, strides=(2, 2), padding='same', activation=activation, kernel_initializer=tf.keras.initializers.Orthogonal, kernel_regularizer=tf.keras.regularizers.l2(reg_constant), name='deconv'+name_suffix)
conv1 = Conv(filters=128, name_suffix='1')(features)
conv2 = Conv(filters=96, name_suffix='2')(conv1)
conv3 = Conv(filters=64, name_suffix='3')(conv2)
conv4 = Conv(filters=32, name_suffix='4')(conv3)
feat = Conv(filters=16, name_suffix='_feat')(conv4)
occ_mask = Conv(filters=1, name_suffix='_occ_mask',
activation=tf.keras.activations.sigmoid)(feat)
if highest_resolution:
return tf.keras.Model(inputs=features, outputs=occ_mask, name='occlusion_estimator'+str(level))
else:
features_up = UpConv(filters=1, name_suffix='_up_feat',
activation=tf.keras.activations.sigmoid)(feat)
occ_mask_up = UpConv(filters=1, name_suffix='_up_occ_mask',
activation=tf.keras.activations.sigmoid)(occ_mask)
return tf.keras.Model(inputs=features, outputs=(occ_mask, features_up, occ_mask_up), name='occlusion_estimator'+str(level))