def __init__(self,
nJoints,
is_training,
batch_size,
img_size=256,
loss_weight_heatmap=1.0,
loss_weight_volume=1.0,
loss_weight_rank=1.0,
loss_weight_2d=1000.0):
self.loss_weight_heatmap = loss_weight_heatmap
self.loss_weight_volume = loss_weight_volume
# Loss weight for no gt training
self.loss_weight_rank = loss_weight_rank
self.loss_weight_2d = loss_weight_2d
self.nJoints = nJoints
self.img_size = img_size
self.is_use_bias = True
self.is_tiny = False
self.is_training = is_training
self.res_utils = mResidualUtils(is_training=self.is_training,
is_use_bias=self.is_use_bias,
is_tiny=self.is_tiny)
self.batch_size = batch_size
self.feature_size = 64
def __init__(self,
nJoints,
is_training,
batch_size,
img_size=256,
coords_scale=1000.0,
coords_2d_scale=255.0,
coords_2d_offset=127.5,
keyp_loss_weight=100.0,
rank_loss_weight=1.0,
coords_loss_weight=1000.0):
self.nJoints = nJoints
self.img_size = img_size
self.is_use_bias = True
self.is_tiny = False
self.is_training = is_training
self.res_utils = mResidualUtils(is_training=self.is_training,
is_use_bias=self.is_use_bias,
is_tiny=self.is_tiny)
self.batch_size = batch_size
self.coords_scale = coords_scale
self.coords_2d_scale = coords_2d_scale
self.coords_2d_offset = coords_2d_offset
self.coords_loss_weight = coords_loss_weight
self.keyp_loss_weight = keyp_loss_weight
self.rank_loss_weight = rank_loss_weight
def __init__(self,
nJoints,
is_training,
batch_size,
img_size=256,
loss_weight_heatmap=1.0,
loss_weight_xyzmap=1.0,
joints_2d_scale=4.0,
joints_3d_scale=1000.0):
self.loss_weight_heatmap = loss_weight_heatmap
self.loss_weight_xyzmap = loss_weight_xyzmap
self.joints_2d_scale = joints_2d_scale
self.joints_3d_scale = joints_3d_scale
self.nJoints = nJoints
self.img_size = img_size
self.is_use_bias = True
self.is_tiny = False
self.is_training = is_training
self.res_utils = mResidualUtils(is_training=self.is_training,
is_use_bias=self.is_use_bias,
is_tiny=self.is_tiny)
self.batch_size = batch_size
self.feature_size = 64
def build_hourglass(inputs,
nOut=256,
nPooling=4,
name='hourglass',
is_training=True,
res_utils=None):
if res_utils is None:
print("Use the default resblock settings!")
res_utils = mResidualUtils(is_training=is_training,
is_tiny=False,
is_use_bias=True)
with tf.variable_scope(name):
# encoding block
up1 = inputs
with tf.variable_scope("up_1"):
for i in range(3):
up1 = res_utils.residual_block(up1,
nOut,
name="res{}".format(i))
with tf.variable_scope("low_1"):
low1 = tf.layers.max_pooling2d(inputs,
pool_size=2,
strides=2,
padding="VALID",
name="down_samplint")
for i in range(3):
low1 = res_utils.residual_block(low1,
nOut,
name="res{}".format(i))
if nPooling > 1:
low2 = build_hourglass(low1,
nOut,
nPooling - 1,
name="inner_hg",
res_utils=res_utils)
else:
# TODO The 2017 version in https://github.com/geopavlakos/c2f-vol-train/blob/master/src/models/hg-stacked-no-int.lua
# contains only one residual block, but the paper contains three
low2 = low1
with tf.variable_scope("mid"):
for i in range(3):
low2 = res_utils.residual_block(low2,
nOut,
name="res{}".format(i))
with tf.variable_scope("low_2"):
low3 = res_utils.residual_block(low2, nOut, name="res")
with tf.variable_scope("up_2"):
cur_shape = low3.get_shape()[1:3].as_list()
up2 = tf.image.resize_nearest_neighbor(
low3, [cur_shape[0] * 2, cur_shape[1] * 2], name="up_sampling")
return tf.add_n([up1, up2], name="out_hg")
def __init__(self,
nJoints,
is_training,
batch_size,
img_size=256,
depth_scale=1000.0,
loss_weight=1000):
self.nJoints = nJoints
self.img_size = img_size
self.is_use_bias = True
self.is_tiny = False
self.is_training = is_training
self.res_utils = mResidualUtils(is_training=self.is_training,
is_use_bias=self.is_use_bias,
is_tiny=self.is_tiny)
self.batch_size = batch_size
self.depth_scale = depth_scale
self.loss_weight = loss_weight
def __init__(self,
nJoints,
is_training,
batch_size,
img_size=256,
loss_weight_volume=1.0,
rank_loss_weight=1.0,
hm_loss_weight=100.0):
self.loss_weight_volume = loss_weight_volume
self.nJoints = nJoints
self.img_size = img_size
self.is_use_bias = True
self.is_tiny = False
self.is_training = is_training
self.res_utils = mResidualUtils(is_training=self.is_training,
is_use_bias=self.is_use_bias,
is_tiny=self.is_tiny)
self.batch_size = batch_size
self.feature_size = 64
self.rank_loss_weight = rank_loss_weight
self.hm_loss_weight = hm_loss_weight