def _upsampling_block(self, bottom, skip_connection, input_channels, output_channels, skip_input_channels, name='upsample', reuse=False):
with tf.variable_scope(name, reuse=reuse):
self._add_to_layers(name + '/deconv', sharedLayers.conv2d_transpose(
bottom, [4, 4, output_channels, input_channels], strides=2, name='deconv'))
self._add_to_layers(name + '/predict', sharedLayers.conv2d(bottom, [
3, 3, input_channels, 1], strides=1, activation=lambda x: x, name='predict'))
self._disparities.append(self._make_disp(self._layers[name + '/predict']))
self._add_to_layers(name + '/up_predict', sharedLayers.conv2d_transpose(self._get_layer_as_input(
name + '/predict'), [4, 4, 1, 1], strides=2, activation=lambda x: x, name='up_predict'))
with tf.variable_scope('join_skip'):
concat_inputs = tf.concat([skip_connection, self._get_layer_as_input(name + '/deconv'), self._get_layer_as_input(name + '/up_predict')], axis=3)
self._add_to_layers(name + '/concat', sharedLayers.conv2d(concat_inputs, [
3, 3, output_channels + skip_input_channels + 1, output_channels], strides=1, activation=lambda x: x, name='concat'))
def _stereo_estimator(self, costs, upsampled_disp=None, scope='fgc-volume-filtering'):
activation = self._leaky_relu()
with tf.variable_scope(scope):
# create initial cost volume
if upsampled_disp is not None:
volume = tf.concat([costs, upsampled_disp], -1)
else:
volume = costs
names = []
# disp-1
names.append('{}/disp1'.format(scope))
input_layer = volume
self._add_to_layers(names[-1], sharedLayers.conv2d(input_layer, [3, 3, volume.get_shape(
).as_list()[3], 128], name='disp-1', bName='biases', activation=activation))
# disp-2:
names.append('{}/disp2'.format(scope))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(names[-1], sharedLayers.conv2d(input_layer, [
3, 3, 128, 128], name='disp-2', bName='biases', activation=activation))
# disp-3
names.append('{}/disp3'.format(scope))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(names[-1], sharedLayers.conv2d(input_layer, [
3, 3, 128, 96], name='disp-3', bName='biases', activation=activation))
# disp-4
names.append('{}/disp4'.format(scope))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(names[-1], sharedLayers.conv2d(input_layer, [
3, 3, 96, 64], name='disp-4', bName='biases', activation=activation))
# disp-5
names.append('{}/disp5'.format(scope))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(names[-1], sharedLayers.conv2d(input_layer, [
3, 3, 64, 32], name='disp-5', bName='biases', activation=activation))
# disp-6
names.append('{}/disp6'.format(scope))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(names[-1], sharedLayers.conv2d(input_layer, [
3, 3, 32, 1], name='disp-6', bName='biases', activation=lambda x: x))
# in_channels = self._get_layer_as_input(names[-1]).get_shape()[-1]
# return sharedLayers.EA(self._get_layer_as_input(names[-1]), self._get_layer_as_input(names[-1]), in_channels, 50, 5, 50)
return self._get_layer_as_input(names[-1])
def _build_network(self, args):
if args['correlation']:
self._add_to_layers('conv1a', sharedLayers.conv2d(
self._left_input_batch, [7, 7, 3, 64], strides=2, name='conv1'))
self._add_to_layers('conv1b', sharedLayers.conv2d(self._right_input_batch, [
7, 7, 3, 64], strides=2, name='conv1', reuse=True))
self._add_to_layers('conv2a', sharedLayers.conv2d(
self._get_layer_as_input('conv1a'), [5, 5, 64, 128], strides=2, name='conv2'))
self._add_to_layers('conv2b', sharedLayers.conv2d(self._get_layer_as_input(
'conv1b'), [5, 5, 64, 128], strides=2, name='conv2', reuse=True))
self._add_to_layers('conv_redir', sharedLayers.conv2d(self._get_layer_as_input(
'conv2a'), [1, 1, 128, 64], strides=1, name='conv_redir'))
self._add_to_layers('corr', sharedLayers.correlation(self._get_layer_as_input(
'conv2a'), self._get_layer_as_input('conv2b'), max_disp=MAX_DISP))
self._add_to_layers('conv3', sharedLayers.conv2d(tf.concat([self._get_layer_as_input('corr'), self._get_layer_as_input(
'conv_redir')], axis=3), [5, 5, MAX_DISP * 2 + 1 + 64, 256], strides=2, name='conv3'))
else:
concat_inputs = tf.concat(
[self._left_img_batch, self._right_input_batch], axis=-1)
self._add_to_layers('conv1', sharedLayers.conv2d(
concat_inputs, [7, 7, 6, 64], strides=2, name='conv1'))
self._add_to_layers('conv2', sharedLayers.conv2d(
self._get_layer_as_input('conv1'), [5, 5, 64, 128], strides=2, name='conv2'))
self._add_to_layers('conv3', sharedLayers.conv2d(
self._get_layer_as_input('conv2'), [5, 5, 128, 256], strides=2, name='conv3'))
self._add_to_layers('conv3/1', sharedLayers.conv2d(
self._get_layer_as_input('conv3'), [3, 3, 256, 256], strides=1, name='conv3/1'))
self._add_to_layers('conv4', sharedLayers.conv2d(self._get_layer_as_input(
'conv3/1'), [3, 3, 256, 512], strides=2, name='conv4'))
self._add_to_layers('conv4/1', sharedLayers.conv2d(
self._get_layer_as_input('conv4'), [3, 3, 512, 512], strides=1, name='conv4/1'))
self._add_to_layers('conv5', sharedLayers.conv2d(self._get_layer_as_input(
'conv4/1'), [3, 3, 512, 512], strides=2, name='conv5'))
self._add_to_layers('conv5/1', sharedLayers.conv2d(
self._get_layer_as_input('conv5'), [3, 3, 512, 512], strides=1, name='conv5/1'))
self._add_to_layers('conv6', sharedLayers.conv2d(self._get_layer_as_input(
'conv5/1'), [3, 3, 512, 1024], strides=2, name='conv6'))
self._add_to_layers('conv6/1', sharedLayers.conv2d(self._get_layer_as_input(
'conv6'), [3, 3, 1024, 1024], strides=1, name='conv6/1'))
self._upsampling_block(self._get_layer_as_input(
'conv6/1'), self._get_layer_as_input('conv5/1'), 1024, 512, 512, name='up5')
self._upsampling_block(self._get_layer_as_input(
'up5/concat'), self._get_layer_as_input('conv4/1'), 512, 256, 512, name='up4')
self._upsampling_block(self._get_layer_as_input(
'up4/concat'), self._get_layer_as_input('conv3/1'), 256, 128, 256, name='up3')
if args['correlation']:
self._upsampling_block(self._get_layer_as_input(
'up3/concat'), self._get_layer_as_input('conv2a'), 128, 64, 128, name='up2')
else:
self._upsampling_block(self._get_layer_as_input(
'up3/concat'), self._get_layer_as_input('conv2'), 128, 64, 128, name='up2')
if args['correlation']:
self._upsampling_block(self._get_layer_as_input(
'up2/concat'), self._get_layer_as_input('conv1a'), 64, 32, 64, name='up1')
else:
self._upsampling_block(self._get_layer_as_input(
'up2/concat'), self._get_layer_as_input('conv1'), 64, 32, 64, name='up1')
self._add_to_layers('prediction', sharedLayers.conv2d(self._get_layer_as_input(
'up1/concat'), [3, 3, 32, 1], strides=1, activation=lambda x: x, name='prediction'))
self._disparities.append(self._make_disp(self._layers['prediction'],2))
############ LOOK BELOW IF DISPNET GIVES WRONG RESULTS #########################
# rescaled_prediction = -preprocessing.rescale_image(self._layers['prediction'], tf.shape(self._left_input_batch)[1:3])
################################################################################
rescaled_prediction = tf.image.resize_images(self._layers['prediction'], tf.shape(self._left_input_batch)[1:3]) * 2
self._layers['rescaled_prediction'] = tf.image.resize_image_with_crop_or_pad(rescaled_prediction, self._restore_shape[0], self._restore_shape[1])
self._disparities.append(self._layers['rescaled_prediction'])
def _pyramid_features(self,
input_batch,
scope='pyramid',
reuse=False,
layer_prefix='pyramid'):
with tf.variable_scope(scope, reuse=reuse):
names = []
activation = self._leaky_relu()
# conv1
names.append('{}/conv1'.format(layer_prefix))
input_layer = input_batch
self._add_to_layers(
names[-1],
sharedLayers.conv2d(
input_layer,
[3, 3, input_batch.get_shape()[-1].value, 16],
strides=2,
name='conv1',
bName='biases',
activation=activation))
# conv2
names.append('{}/conv2'.format(layer_prefix))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(
names[-1],
sharedLayers.conv2d(input_layer, [3, 3, 16, 16],
strides=1,
name='conv2',
bName='biases',
activation=activation))
# conv3
names.append('{}/conv3'.format(layer_prefix))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(
names[-1],
sharedLayers.conv2d(input_layer, [3, 3, 16, 32],
strides=2,
name='conv3',
bName='biases',
activation=activation))
# conv4
names.append('{}/conv4'.format(layer_prefix))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(
names[-1],
sharedLayers.conv2d(input_layer, [3, 3, 32, 32],
strides=1,
name='conv4',
bName='biases',
activation=activation))
# conv5
names.append('{}/conv5'.format(layer_prefix))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(
names[-1],
sharedLayers.conv2d(input_layer, [3, 3, 32, 64],
strides=2,
name='conv5',
bName='biases',
activation=activation))
# conv6
names.append('{}/conv6'.format(layer_prefix))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(
names[-1],
sharedLayers.conv2d(input_layer, [3, 3, 64, 64],
strides=1,
name='conv6',
bName='biases',
activation=activation))
# conv7
names.append('{}/conv7'.format(layer_prefix))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(
names[-1],
sharedLayers.conv2d(input_layer, [3, 3, 64, 96],
strides=2,
name='conv7',
bName='biases',
activation=activation))
# conv8
names.append('{}/conv8'.format(layer_prefix))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(
names[-1],
sharedLayers.conv2d(input_layer, [3, 3, 96, 96],
strides=1,
name='conv8',
bName='biases',
activation=activation))
# conv9
names.append('{}/conv9'.format(layer_prefix))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(
names[-1],
sharedLayers.conv2d(input_layer, [3, 3, 96, 128],
strides=2,
name='conv9',
bName='biases',
activation=activation))
# conv10
names.append('{}/conv10'.format(layer_prefix))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(
names[-1],
sharedLayers.conv2d(input_layer, [3, 3, 128, 128],
strides=1,
name='conv10',
bName='biases',
activation=activation))
# conv11
names.append('{}/conv11'.format(layer_prefix))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(
names[-1],
sharedLayers.conv2d(input_layer, [3, 3, 128, 192],
strides=2,
name='conv11',
bName='biases',
activation=activation))
# conv12
names.append('{}/conv12'.format(layer_prefix))
input_layer = self._get_layer_as_input(names[-2])
self._add_to_layers(
names[-1],
sharedLayers.conv2d(input_layer, [3, 3, 192, 192],
strides=1,
name='conv12',
bName='biases',
activation=activation))