def __init__(self, scope=None):
with tf.variable_scope(scope, reuse=True):
colour_channels = 1 if opt.grey_scale else 3
input_uint8_1 = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_1')
input_uint8_1r = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_1r')
input_uint8_2 = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_2')
input_uint8_2r = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_2r')
input_intrinsic = tf.placeholder(tf.float32, [3, 3])
input_1 = preprocess_image(input_uint8_1)
input_2 = preprocess_image(input_uint8_2)
input_1r = preprocess_image(input_uint8_1r)
input_2r = preprocess_image(input_uint8_2r)
feature1_disp = feature_pyramid_disp(input_1, reuse=True)
feature1r_disp = feature_pyramid_disp(input_1r, reuse=True)
pred_disp = disp_godard(
input_1,
input_1r,
feature1_disp,
feature1r_disp,
opt,
is_training=False)
self.input_1 = input_uint8_1
self.input_2 = input_uint8_2
self.input_r = input_uint8_1r
self.input_2r = input_uint8_2r
self.input_intrinsic = input_intrinsic
self.pred_disp = pred_disp[0][:, :, :, 0:1]
# Placeholder created for interface consistency
self.pred_flow_optical = tf.constant(0.0)
self.pred_flow_rigid = tf.constant(0.0)
self.pred_disp2 = tf.constant(0.0)
self.pred_mask = tf.constant(0.0)
def __init__(self, scope=None):
with tf.variable_scope(scope, reuse=True):
colour_channels = 1 if opt.grey_scale else 3
input_uint8_1 = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_1')
input_uint8_1r = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_1r')
input_uint8_2 = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_2')
input_uint8_2r = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_2r')
input_intrinsic = tf.placeholder(tf.float32, [3, 3])
cam2pix, pix2cam = get_multi_scale_intrinsics(input_intrinsic,
opt.num_scales)
cam2pix = tf.expand_dims(cam2pix, axis=0)
pix2cam = tf.expand_dims(pix2cam, axis=0)
input_1 = preprocess_image(input_uint8_1)
input_2 = preprocess_image(input_uint8_2)
input_1r = preprocess_image(input_uint8_1r)
input_2r = preprocess_image(input_uint8_2r)
feature1 = feature_pyramid_flow(input_1, reuse=True)
feature2 = feature_pyramid_flow(input_2, reuse=True)
optical_flows = construct_model_pwc_full(input_1, input_2,
feature1, feature2)
self.input_1 = input_uint8_1
self.input_2 = input_uint8_2
self.input_r = input_uint8_1r
self.input_2r = input_uint8_2r
self.input_intrinsic = input_intrinsic
self.pred_flow_optical = optical_flows[0]
# Placeholder created for interface consistency
self.pred_flow_rigid = tf.constant(0.0)
self.pred_disp = tf.constant(0.0)
self.pred_disp2 = tf.constant(0.0)
self.pred_mask = tf.constant(0.0)
def __init__(self, scope=None):
with tf.variable_scope(scope, reuse=True):
colour_channels = 1 if opt.grey_scale else 3
input_uint8_1 = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_1')
input_uint8_1r = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_1r')
input_uint8_2 = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_2')
input_uint8_2r = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_2r')
input_intrinsic = tf.placeholder(tf.float32, [3, 3])
cam2pix, pix2cam = get_multi_scale_intrinsics(input_intrinsic,
opt.num_scales)
cam2pix = tf.expand_dims(cam2pix, axis=0)
pix2cam = tf.expand_dims(pix2cam, axis=0)
input_1 = preprocess_image(input_uint8_1)
input_2 = preprocess_image(input_uint8_2)
input_1r = preprocess_image(input_uint8_1r)
input_2r = preprocess_image(input_uint8_2r)
feature1_disp = feature_pyramid_disp(input_1, reuse=True)
feature1r_disp = feature_pyramid_disp(input_1r, reuse=True)
feature2_disp = feature_pyramid_disp(input_2, reuse=True)
feature2r_disp = feature_pyramid_disp(input_2r, reuse=True)
feature1_flow = feature_pyramid_flow(input_1, reuse=True)
feature2_flow = feature_pyramid_flow(input_2, reuse=True)
pred_disp = disp_godard(
input_1,
input_1r,
feature1_disp,
feature1r_disp,
opt,
is_training=False)
pred_disp_rev = disp_godard(
input_2,
input_2r,
feature2_disp,
feature2r_disp,
opt,
is_training=False)
pred_poses = pose_exp_net(input_1, input_2)
optical_flows = construct_model_pwc_full(
input_1, input_2, feature1_flow, feature2_flow)
optical_flows_rev = construct_model_pwc_full(
input_2, input_1, feature2_flow, feature1_flow)
s = 0
occu_mask = tf.clip_by_value(
transformerFwd(
tf.ones(
shape=[
1, opt.img_height // (2**s),
opt.img_width // (2**s), 1
],
dtype='float32'),
optical_flows_rev[s],
[opt.img_height // (2**s), opt.img_width // (2**s)]),
clip_value_min=0.0,
clip_value_max=1.0)
depth_flow, pose_mat, disp1_trans, small_mask = inverse_warp_new(
1.0 / pred_disp[0][:, :, :, 0:1],
1.0 / pred_disp_rev[0][:, :, :, 0:1], pred_poses,
cam2pix[:, 0, :, :], pix2cam[:, 0, :, :], optical_flows[0],
occu_mask)
flow_diff = tf.sqrt(
tf.reduce_sum(
tf.square(depth_flow - optical_flows[0]),
axis=3,
keep_dims=True))
flow_diff_mask = tf.cast(flow_diff < (opt.flow_diff_threshold),
tf.float32)
occu_region = tf.cast(occu_mask < 0.5, tf.float32)
ref_exp_mask = tf.clip_by_value(
flow_diff_mask + occu_region,
clip_value_min=0.0,
clip_value_max=1.0)
self.input_1 = input_uint8_1
self.input_2 = input_uint8_2
self.input_r = input_uint8_1r
self.input_2r = input_uint8_2r
self.input_intrinsic = input_intrinsic
self.pred_pose_mat = pose_mat[0, :, :]
self.pred_flow_rigid = depth_flow
self.pred_flow_optical = optical_flows[0]
self.pred_disp = pred_disp[0][:, :, :, 0:1]
self.pred_disp2 = disp1_trans*0.0 + \
transformer_old(pred_disp_rev[0][:,:,:,0:1], optical_flows[0], [opt.img_height, opt.img_width])*(1.0-0.0)
self.pred_mask = 1.0 - ref_exp_mask
def __init__(self, scope=None):
with tf.variable_scope(scope, reuse=True):
colour_channels = 1 if opt.grey_scale else 3
input_uint8_1 = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_1')
input_uint8_1r = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_1r')
input_uint8_2 = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_2')
input_uint8_2r = tf.placeholder(
tf.uint8, [1, opt.img_height, opt.img_width, colour_channels],
name='raw_input_2r')
input_intrinsic = tf.placeholder(tf.float32, [3, 3])
cam2pix, pix2cam = get_multi_scale_intrinsics(input_intrinsic,
opt.num_scales)
cam2pix = tf.expand_dims(cam2pix, axis=0)
pix2cam = tf.expand_dims(pix2cam, axis=0)
input_1 = preprocess_image(input_uint8_1)
input_2 = preprocess_image(input_uint8_2)
input_1r = preprocess_image(input_uint8_1r)
input_2r = preprocess_image(input_uint8_2r)
feature1_disp = feature_pyramid_disp(input_1, reuse=True)
feature1r_disp = feature_pyramid_disp(input_1r, reuse=True)
feature1_flow = feature_pyramid_flow(input_1, reuse=True)
feature2_flow = feature_pyramid_flow(input_2, reuse=True)
pred_disp = disp_godard(
input_1,
input_1r,
feature1_disp,
feature1r_disp,
opt,
is_training=False)
pred_poses = pose_exp_net(input_1, input_2)
optical_flows = construct_model_pwc_full(
input_1, input_2, feature1_flow, feature2_flow)
s = 0
depth_flow, pose_mat = inverse_warp(
1.0 / pred_disp[s][:, :, :, 0:1],
pred_poses,
cam2pix[:, s, :, :], ## [batchsize, scale, 3, 3]
pix2cam[:, s, :, :])
self.input_1 = input_uint8_1
self.input_2 = input_uint8_2
self.input_r = input_uint8_1r
self.input_2r = input_uint8_2r
self.input_intrinsic = input_intrinsic
self.pred_flow_rigid = depth_flow
self.pred_flow_optical = optical_flows[0]
self.pred_disp = pred_disp[0][:, :, :, 0:1]
self.pred_pose_mat = pose_mat[0, :, :]
# Placeholder created for interface consistency
self.pred_disp2 = tf.constant(0.0)
self.pred_mask = tf.constant(0.0)