def _conf_slice_refinement_net(conf_slice, img_features, is_training):
"""
Small network with shared weigts leearning to refine the confidence volume using the input texture.
Uses shared weigts for all.
"""
h = int(conf_slice.shape[1])
w = int(conf_slice.shape[2])
#with tf.variable_scope("conf_slice_refinement_size_%d_%d"%(h,w), reuse=tf.AUTO_REUSE):
with tf.variable_scope("conf_slice_refinement", reuse=tf.AUTO_REUSE):
conf_slice_out = conv2d(inputs=conf_slice, filters=16, kernel_size=3, name = 'refine_conf_conv1' )
conf_slice_out = tf.layers.batch_normalization(inputs=conf_slice_out, training=is_training, name='refine_conf_bn1' )
conf_slice_out = tf.nn.leaky_relu(features=conf_slice_out, name='refine_conf_leaky1')
conf_slice_out = resnet_block(inputs=conf_slice_out, filters=16, kernel_size=3, dilation_rate=1,name='refine_conf_res1')
conf_slice_out = resnet_block(inputs=conf_slice_out, filters=16, kernel_size=3, dilation_rate=2, name='refine_conf_res2')
#Resize image features
image_out = tf.image.resize_bilinear(img_features, conf_slice.shape[1:3])
#
concat_out = tf.concat([image_out,conf_slice_out],axis=3,name='refine_concat')
#
concat_out = resnet_block(inputs=concat_out, filters=32, kernel_size=3, dilation_rate=4, name='refine_concat_res0')
#concat_out = resnet_block(inputs=concat_out, filters=32, kernel_size=3, dilation_rate=8, name='refine_concat_res1')
#concat_out = resnet_block(inputs=concat_out, filters=32, kernel_size=3, dilation_rate=1, name='refine_concat_res2')
#concat_out = resnet_block(inputs=concat_out, filters=32, kernel_size=3, dilation_rate=1, name='refine_concat_res3')
#note can use tanh
conf_residual = conv2d(inputs=concat_out, filters=1, kernel_size=3, name = 'refine_concat_conv1')
conf_residual = tf.nn.tanh(conf_residual, name='conf_residual_tanh')
#
refined_conf_slice = tf.clip_by_value(conf_slice + conf_residual, 0, 1)
return refined_conf_slice
def _tower_feature_net(img, is_training):
"""
Definition for the feature network.
img: the input tensor for a batch of images
Return a tensor of features.
"""
with tf.variable_scope("feature_tower", reuse=tf.AUTO_REUSE):
out = conv2d(inputs=img, filters=32, kernel_size=3, name='tower_conv1')
out = resnet_block(inputs=out,
filters=32,
kernel_size=3,
name='tower_res1')
out = resnet_block(inputs=out,
filters=32,
kernel_size=3,
name='tower_res2')
out = resnet_block(inputs=out,
filters=32,
kernel_size=3,
name='tower_res3')
out = conv2d(inputs=out,
filters=32,
kernel_size=3,
strides=2,
name='tower_conv2')
out = tf.layers.batch_normalization(inputs=out,
training=is_training,
name='tower_bn1')
out = tf.nn.leaky_relu(features=out, name='tower_leaky1')
out = conv2d(inputs=out,
filters=32,
kernel_size=3,
strides=2,
name='tower_conv3')
out = tf.layers.batch_normalization(inputs=out,
training=is_training,
name='tower_bn2')
out = tf.nn.leaky_relu(features=out, name='tower_leaky2')
out = conv2d(inputs=out,
filters=32,
kernel_size=3,
strides=2,
name='tower_conv4')
out = tf.layers.batch_normalization(inputs=out,
training=is_training,
name='tower_bn3')
tower_feature = tf.nn.leaky_relu(features=out, name='tower_leaky3')
return tower_feature
def _image_feature_net(img,is_training):
with tf.variable_scope("image_features"):
image_out = conv2d(inputs=img, filters=16, kernel_size=3, name = 'refine_image_conv1')
image_out = tf.layers.batch_normalization(inputs=image_out, training=is_training, name='refine_image_bn1')
image_out = tf.nn.leaky_relu(features=image_out, name='refine_image_leaky1')
image_out = resnet_block(inputs=image_out, filters=16, kernel_size=3, dilation_rate=1,name='refine_image_res1')
image_out = resnet_block(inputs=image_out, filters=16, kernel_size=3, dilation_rate=2, name='refine_image_res2')
return image_out
def _blur_image_refinement_net(img,
upsampled_blur_image,
scale_name="1",
is_training=False):
"""
Refines the upsampled blur image using the AIF image.
This is the same block as stereonet.
"""
with tf.variable_scope("guided_upsampling_scale_" + scale_name):
#get features from the upsampled blur image
blur_image_features = conv2d(inputs=upsampled_blur_image,
filters=16,
kernel_size=3,
name='refine_blur_image_conv1_' +
scale_name)
blur_image_features = tf.layers.batch_normalization(
inputs=blur_image_features,
training=is_training,
name='refine_blur_image_bn1_' + scale_name)
blur_image_features = tf.nn.leaky_relu(
features=blur_image_features,
name='refine_blur_image_leaky1_' + scale_name)
blur_image_features = resnet_block(inputs=blur_image_features,
filters=16,
kernel_size=3,
dilation_rate=1,
name='refine_blur_image_res1_' +
scale_name)
blur_image_features = resnet_block(inputs=blur_image_features,
filters=16,
kernel_size=3,
dilation_rate=2,
name='refine_blur_image_res2_' +
scale_name)
#get features from the AIF image (downsapled b4 if needed)
aif_image_features = conv2d(inputs=img,
filters=16,
kernel_size=3,
name='refine_image_conv1_' + scale_name)
aif_image_features = tf.layers.batch_normalization(
inputs=aif_image_features,
training=is_training,
name='refine_image_bn1_' + scale_name)
aif_image_features = tf.nn.leaky_relu(features=aif_image_features,
name='refine_image_leaky1_' +
scale_name)
aif_image_features = resnet_block(inputs=aif_image_features,
filters=16,
kernel_size=3,
dilation_rate=1,
name='refine_image_res1_' +
scale_name)
aif_image_features = resnet_block(inputs=aif_image_features,
filters=16,
kernel_size=3,
dilation_rate=2,
name='refine_image_res2_' +
scale_name)
#cat
concat_out = tf.concat([aif_image_features, blur_image_features],
axis=3,
name='refine_concat_' + scale_name)
#some seridual blocks
concat_out = resnet_block(inputs=concat_out,
filters=32,
kernel_size=3,
dilation_rate=4,
name='refine_concat_res0_' + scale_name)
concat_out = resnet_block(inputs=concat_out,
filters=32,
kernel_size=3,
dilation_rate=8,
name='refine_concat_res1_' + scale_name)
concat_out = resnet_block(inputs=concat_out,
filters=32,
kernel_size=3,
dilation_rate=1,
name='refine_concat_res2_' + scale_name)
concat_out = resnet_block(inputs=concat_out,
filters=32,
kernel_size=3,
dilation_rate=1,
name='refine_concat_res3_' + scale_name)
#residual to add to the refocus image
blur_image_residual = conv2d(inputs=concat_out,
filters=3,
kernel_size=3,
name='refine_concat_conv1_' + scale_name)
#normalise it
blur_image_residual = tf.nn.tanh(blur_image_residual)
#add it and return (note: normalised)
high_res_blur_image = upsampled_blur_image + blur_image_residual
#clipping?
#high_res_blur_image = tf.clip_by_value(high_res_blur_image, 0, 1)
return high_res_blur_image
def _disparity_map_refinement_net(img, upsampled_disparity_map, scale_name,
is_training):
"""
Step to upsample and refine a disparity map, aided by the input image.
"""
with tf.variable_scope("guided_upsampling_scale_" + scale_name):
#
disparity_out = conv2d(inputs=upsampled_disparity_map,
filters=16,
kernel_size=3,
name='refine_disparity_conv1_' + scale_name)
disparity_out = tf.layers.batch_normalization(
inputs=disparity_out,
training=is_training,
name='refine_disparity_bn1_' + scale_name)
disparity_out = tf.nn.leaky_relu(features=disparity_out,
name='refine_disparity_leaky1_' +
scale_name)
disparity_out = resnet_block(inputs=disparity_out,
filters=16,
kernel_size=3,
dilation_rate=1,
name='refine_disparity_res1_' +
scale_name)
disparity_out = resnet_block(inputs=disparity_out,
filters=16,
kernel_size=3,
dilation_rate=2,
name='refine_disparity_res2_' +
scale_name)
#
image_out = conv2d(inputs=img,
filters=16,
kernel_size=3,
name='refine_image_conv1_' + scale_name)
image_out = tf.layers.batch_normalization(inputs=image_out,
training=is_training,
name='refine_image_bn1_' +
scale_name)
image_out = tf.nn.leaky_relu(features=image_out,
name='refine_image_leaky1_' + scale_name)
image_out = resnet_block(inputs=image_out,
filters=16,
kernel_size=3,
dilation_rate=1,
name='refine_image_res1_' + scale_name)
image_out = resnet_block(inputs=image_out,
filters=16,
kernel_size=3,
dilation_rate=2,
name='refine_image_res2_' + scale_name)
#
concat_out = tf.concat([image_out, disparity_out],
axis=3,
name='refine_concat_' + scale_name)
#
concat_out = resnet_block(inputs=concat_out,
filters=32,
kernel_size=3,
dilation_rate=4,
name='refine_concat_res0_' + scale_name)
concat_out = resnet_block(inputs=concat_out,
filters=32,
kernel_size=3,
dilation_rate=8,
name='refine_concat_res1_' + scale_name)
concat_out = resnet_block(inputs=concat_out,
filters=32,
kernel_size=3,
dilation_rate=1,
name='refine_concat_res2_' + scale_name)
concat_out = resnet_block(inputs=concat_out,
filters=32,
kernel_size=3,
dilation_rate=1,
name='refine_concat_res3_' + scale_name)
#
disparity_residual = conv2d(inputs=concat_out,
filters=1,
kernel_size=3,
name='refine_concat_conv1_' + scale_name)
#
high_res_disparity_map = upsampled_disparity_map + disparity_residual
return high_res_disparity_map