def PRRB(xyz,x_points,is_training, bn_decay, i):
c=x_points.shape[2]
# (B,N,512)
x1 = tf_util.conv1d(x_points, c, 1, padding='VALID', bn=True,
is_training=is_training, scope='fc_URB_'+str(i),bn_decay=bn_decay,activation_fn=None)
x11=x1
# (B,N,512)
_, x11, _ = pointSIFT_module(xyz, x11, radius=0.2, out_channel=c, is_training=is_training,
bn_decay=bn_decay,bn=True, scope='fsift_URB_'+str(i)+'_1')
x11 = lrelu(x11, 0)
# (B,N,512)
_, x11, _ = pointSIFT_module(xyz, x11, radius=0.2, out_channel=c, is_training=is_training,
bn_decay=None, bn=False, scope='fsiftnobn_URB_'+str(i)+'_1')
# (B,N,512)
x=tf.add(x11,x1)
x = lrelu(x, 0)
return x
def get_model(point_cloud, is_training, num_class, bn_decay=None):
""" Semantic segmentation PointNet, input is BxNx5, output Bxnum_class """
batch_size = point_cloud.get_shape()[0].value
num_point = point_cloud.get_shape()[1].value
end_points = {}
l0_xyz = tf.slice(point_cloud, [0, 0, 0], [-1, -1, 3]) # x, y ,z => BxNx3
l0_points = tf.slice(point_cloud, [0, 0, 3], [-1, -1, 2]) # 强度，返回值 =>BxNx2
end_points['l0_xyz'] = l0_xyz
# c0
c0_l0_xyz, c0_l0_points, c0_l0_indices = pointSIFT_res_module(
l0_xyz,
l0_points,
radius=0.1,
out_channel=64,
is_training=is_training,
bn_decay=bn_decay,
scope='layer0_c0',
merge='concat')
l1_xyz, l1_points, l1_indices = pointnet_sa_module_layer1(
c0_l0_xyz,
c0_l0_points,
npoint=512,
radius=0.1,
nsample=32,
mlp=[64, 128],
mlp2=None,
group_all=False,
is_training=is_training,
bn_decay=bn_decay,
scope='layer1')
# c1
c0_l1_xyz, c0_l1_points, c0_l1_indices = pointSIFT_res_module(
l1_xyz,
l1_points,
radius=0.25,
out_channel=128,
is_training=is_training,
bn_decay=bn_decay,
scope='layer1_c0')
l2_xyz, l2_points, l2_indices = pointnet_sa_module(c0_l1_xyz,
c0_l1_points,
npoint=256,
radius=0.2,
nsample=32,
mlp=[128, 256],
mlp2=None,
group_all=False,
is_training=is_training,
bn_decay=bn_decay,
scope='layer2')
# c2
c0_l2_xyz, c0_l2_points, c0_l2_indices = pointSIFT_res_module(
l2_xyz,
l2_points,
radius=0.5,
out_channel=256,
is_training=is_training,
bn_decay=bn_decay,
scope='layer2_c0')
c1_l2_xyz, c1_l2_points, c1_l2_indices = pointSIFT_res_module(
c0_l2_xyz,
c0_l2_points,
radius=0.5,
out_channel=512,
is_training=is_training,
bn_decay=bn_decay,
scope='layer2_c1',
same_dim=True)
l2_cat_points = tf.concat([c0_l2_points, c1_l2_points], axis=-1)
fc_l2_points = tf_util.conv1d(l2_cat_points,
512,
1,
padding='VALID',
bn=True,
is_training=is_training,
scope='conv_2_fc',
bn_decay=bn_decay)
# c3
l3_xyz, l3_points, l3_indices = pointnet_sa_module(c1_l2_xyz,
fc_l2_points,
npoint=64,
radius=0.4,
nsample=32,
mlp=[512, 512],
mlp2=None,
group_all=False,
is_training=is_training,
bn_decay=bn_decay,
scope='layer3')
l2_points = pointnet_fp_module(l2_xyz,
l3_xyz,
l2_points,
l3_points, [512, 512],
is_training,
bn_decay,
scope='fa_layer2')
_, l2_points_1, _ = pointSIFT_module(l2_xyz,
l2_points,
radius=0.5,
out_channel=512,
is_training=is_training,
bn_decay=bn_decay,
scope='fa_layer2_c0')
_, l2_points_2, _ = pointSIFT_module(l2_xyz,
l2_points,
radius=0.5,
out_channel=512,
is_training=is_training,
bn_decay=bn_decay,
scope='fa_layer2_c1')
_, l2_points_3, _ = pointSIFT_module(l2_xyz,
l2_points,
radius=0.5,
out_channel=512,
is_training=is_training,
bn_decay=bn_decay,
scope='fa_layer2_c2')
l2_points = tf.concat([l2_points_1, l2_points_2, l2_points_3], axis=-1)
l2_points = tf_util.conv1d(l2_points,
512,
1,
padding='VALID',
bn=True,
is_training=is_training,
scope='fa_2_fc',
bn_decay=bn_decay)
l1_points = pointnet_fp_module(l1_xyz,
l2_xyz,
l1_points,
l2_points, [256, 256],
is_training,
bn_decay,
scope='fa_layer3')
_, l1_points_1, _ = pointSIFT_module(l1_xyz,
l1_points,
radius=0.25,
out_channel=256,
is_training=is_training,
bn_decay=bn_decay,
scope='fa_layer3_c0')
_, l1_points_2, _ = pointSIFT_module(l1_xyz,
l1_points_1,
radius=0.25,
out_channel=256,
is_training=is_training,
bn_decay=bn_decay,
scope='fa_layer3_c1')
l1_points = tf.concat([l1_points_1, l1_points_2], axis=-1)
l1_points = tf_util.conv1d(l1_points,
256,
1,
padding='VALID',
bn=True,
is_training=is_training,
scope='fa_1_fc',
bn_decay=bn_decay)
l0_points = pointnet_fp_module(l0_xyz,
l1_xyz,
l0_points,
l1_points, [128, 128, 128],
is_training,
bn_decay,
scope='fa_layer4')
_, l0_points, _ = pointSIFT_module(l0_xyz,
l0_points,
radius=0.1,
out_channel=128,
is_training=is_training,
bn_decay=bn_decay,
scope='fa_layer4_c0')
# FC layers
net = tf_util.conv1d(l0_points,
128,
1,
padding='VALID',
bn=True,
is_training=is_training,
scope='fc1',
bn_decay=bn_decay)
net = tf_util.dropout(net,
keep_prob=0.5,
is_training=is_training,
scope='dp1')
net = tf_util.conv1d(net,
num_class,
1,
padding='VALID',
activation_fn=None,
scope='fc2')
return net, end_points
def get_model(point_cloud,
is_training,
num_class,
bn_decay=None,
feature=None):
""" Semantic segmentation PointNet, input is B x N x 3, output B x num_class """
end_points = {}
l0_xyz = point_cloud[:, :, :3]
l0_points = point_cloud
end_points['l0_xyz'] = l0_xyz
# c0
c0_l0_xyz, c0_l0_points, c0_l0_indices = pointSIFT_res_module(
l0_xyz,
l0_points,
radius=0.1,
out_channel=64,
is_training=is_training,
bn_decay=bn_decay,
scope='layer0_c0',
merge='concat')
###c0_10_points(B,8192,64)
l1_xyz, l1_points, l1_indices = pointnet_sa_module(c0_l0_xyz,
c0_l0_points,
npoint=1024,
radius=0.1,
nsample=32,
mlp=[64, 128],
mlp2=None,
group_all=False,
is_training=is_training,
bn_decay=bn_decay,
scope='layer1')
###l1_points:(B,1024,128)
# c1
c0_l1_xyz, c0_l1_points, c0_l1_indices = pointSIFT_res_module(
l1_xyz,
l1_points,
radius=0.25,
out_channel=128,
is_training=is_training,
bn_decay=bn_decay,
scope='layer1_c0')
###c0_11_points(B,1024,128)
l2_xyz, l2_points, l2_indices = pointnet_sa_module(c0_l1_xyz,
c0_l1_points,
npoint=256,
radius=0.2,
nsample=32,
mlp=[128, 256],
mlp2=None,
group_all=False,
is_training=is_training,
bn_decay=bn_decay,
scope='layer2')
###l2_points(B,256,256)
# c2
c0_l2_xyz, c0_l2_points, c0_l2_indices = pointSIFT_res_module(
l2_xyz,
l2_points,
radius=0.5,
out_channel=256,
is_training=is_training,
bn_decay=bn_decay,
scope='layer2_c0')
###c0_12_points(B,256,256)
c1_l2_xyz, c1_l2_points, c1_l2_indices = pointSIFT_res_module(
c0_l2_xyz,
c0_l2_points,
radius=0.5,
out_channel=512,
is_training=is_training,
bn_decay=bn_decay,
scope='layer2_c1',
same_dim=True)
###c1_12_points(B,256,512)
l2_cat_points = tf.concat([c0_l2_points, c1_l2_points], axis=-1)
###12_cat_points(B,256,768)
fc_l2_points = tf_util.conv1d(l2_cat_points,
512,
1,
padding='VALID',
bn=True,
is_training=is_training,
scope='conv_2_fc',
bn_decay=bn_decay)
###fc_cat_points(B,256,512)
# c3
l3_xyz, l3_points, l3_indices = pointnet_sa_module(c1_l2_xyz,
fc_l2_points,
npoint=64,
radius=0.4,
nsample=32,
mlp=[512, 512],
mlp2=None,
group_all=False,
is_training=is_training,
bn_decay=bn_decay,
scope='layer3')
###l3_point (B,64,512)
l2_points = pointnet_fp_module(l2_xyz,
l3_xyz,
l2_points,
l3_points, [512, 512],
is_training,
bn_decay,
scope='fa_layer2')
###l2_points(B,256,512)
_, l2_points_1, _ = pointSIFT_module(l2_xyz,
l2_points,
radius=0.5,
out_channel=512,
is_training=is_training,
bn_decay=bn_decay,
scope='fa_layer2_c0')
_, l2_points_2, _ = pointSIFT_module(l2_xyz,
l2_points,
radius=0.5,
out_channel=512,
is_training=is_training,
bn_decay=bn_decay,
scope='fa_layer2_c1')
_, l2_points_3, _ = pointSIFT_module(l2_xyz,
l2_points,
radius=0.5,
out_channel=512,
is_training=is_training,
bn_decay=bn_decay,
scope='fa_layer2_c2')
###l2_points_1,2,3(B,256,512)
l2_points = tf.concat([l2_points_1, l2_points_2, l2_points_3], axis=-1)
###l2_points(B,256,1536)
l2_points = tf_util.conv1d(l2_points,
512,
1,
padding='VALID',
bn=True,
is_training=is_training,
scope='fa_2_fc',
bn_decay=bn_decay)
###l2_points(B,256,512)
l1_points = pointnet_fp_module(l1_xyz,
l2_xyz,
l1_points,
l2_points, [256, 256],
is_training,
bn_decay,
scope='fa_layer3')
###l1_points:(B,1024,256)
_, l1_points_1, _ = pointSIFT_module(l1_xyz,
l1_points,
radius=0.25,
out_channel=256,
is_training=is_training,
bn_decay=bn_decay,
scope='fa_layer3_c0')
###l1_points_1(B,1024,256)
_, l1_points_2, _ = pointSIFT_module(l1_xyz,
l1_points_1,
radius=0.25,
out_channel=256,
is_training=is_training,
bn_decay=bn_decay,
scope='fa_layer3_c1')
###l1_points_1(B,1024,256)
l1_points = tf.concat([l1_points_1, l1_points_2], axis=-1)
###l1_points_1(B,1024,512)
l1_points = tf_util.conv1d(l1_points,
256,
1,
padding='VALID',
bn=True,
is_training=is_training,
scope='fa_1_fc',
bn_decay=bn_decay)
###l1_points_1(B,1024,256)
l0_points = pointnet_fp_module(l0_xyz,
l1_xyz,
l0_points,
l1_points, [128, 128, 128],
is_training,
bn_decay,
scope='fa_layer4')
###l0_points(B,8192, 256)
_, l0_points, _ = pointSIFT_module(l0_xyz,
l0_points,
radius=0.1,
out_channel=128,
is_training=is_training,
bn_decay=bn_decay,
scope='fa_layer4_c0')
###l0_points(B,8192,128)
# FC layers
net = tf_util.conv1d(l0_points,
128,
1,
padding='VALID',
bn=True,
is_training=is_training,
scope='fc1',
bn_decay=bn_decay)
###l0_points(B,8192,128)
net = tf_util.dropout(net,
keep_prob=0.5,
is_training=is_training,
scope='dp1')
net = tf_util.conv1d(net,
num_class,
1,
padding='VALID',
activation_fn=None,
scope='fc2')
return net, end_points
def get_model(point_cloud, is_training, num_class, bn_decay=None, feature=None):
""" Semantic segmentation PointNet, input is B x N x 3, output B x num_class """
end_points = {}
l0_xyz = point_cloud[:,:,:3]
l0_points = point_cloud
end_points['l0_xyz'] = l0_xyz
#################################################### E N C O D E R ######################################################
################################################### c0
c0_l0_xyz, c0_l0_points, c0_l0_indices = pointSIFT_res_module(l0_xyz, l0_points, radius=0.1, out_channel=64,
is_training=is_training, bn_decay=bn_decay,
scope='layer0_c0', merge='concat')
###c0_10_points(B,N,64)
l1_xyz, l1_points, l1_indices = pointnet_sa_module(c0_l0_xyz, c0_l0_points, npoint=4096, radius=0.1, nsample=32,
mlp=[128], mlp2=None, group_all=False,
is_training=is_training, bn_decay=bn_decay, scope='layer0')
###l1_points:(B,4096,128)
################################################### c1
c0_l1_xyz, c0_l1_points, c0_l1_indices = pointSIFT_res_module(l1_xyz, l1_points, radius=0.2, out_channel=128,
is_training=is_training, bn_decay=bn_decay,
scope='layer1_c0',merge='concat')
###c0_11_points(B,4096,128)
l2_xyz, l2_points, l2_indices = pointnet_sa_module(c0_l1_xyz, c0_l1_points, npoint=1024, radius=0.2, nsample=32,
mlp=[256], mlp2=None, group_all=False,
is_training=is_training, bn_decay=bn_decay, scope='layer1')
###l2_points(B,1024,256)
#################################################### c2
c0_l2_xyz, c0_l2_points, c0_l2_indices = pointSIFT_res_module(l2_xyz, l2_points, radius=0.2, out_channel=256,
is_training=is_training, bn_decay=bn_decay,
scope='layer2_c0',merge='concat')
###c0_12_points(B,1024,256)
l3_xyz, l3_points, l3_indices = pointnet_sa_module(c0_l2_xyz, c0_l2_points, npoint=256, radius=0.2, nsample=32,
mlp=[512], mlp2=None, group_all=False,
is_training=is_training, bn_decay=bn_decay, scope='layer2')
#l3_points(B,256,512)
#################################################### c3
c0_l3_xyz, c0_l3_points, c0_l3_indices = pointSIFT_res_module(l3_xyz, l3_points, radius=0.2, out_channel=512,
is_training=is_training, bn_decay=bn_decay,
scope='layer3_c0',merge='concat')
###c0_13_points(B,256,512)
l4_xyz, l4_points, l4_indices = pointnet_sa_module(c0_l3_xyz, c0_l3_points, npoint=64, radius=0.2, nsample=32,
mlp=[1024], mlp2=None, group_all=False,
is_training=is_training, bn_decay=bn_decay, scope='layer3')
# l4_points(B,64,1024)
#################################################### c4
c0_l4_xyz, c0_l4_points, c0_l4_indices = pointSIFT_res_module(l4_xyz, l4_points, radius=0.2, out_channel=1024,
is_training=is_training, bn_decay=bn_decay,
scope='layer4_c0',merge='concat')
###c0_14_points(B,64,1024)
l5_xyz, l5_points, l5_indices = pointnet_sa_module(c0_l4_xyz, c0_l4_points, npoint=16, radius=0.2, nsample=32,
mlp=[2048], mlp2=None, group_all=False,
is_training=is_training, bn_decay=bn_decay, scope='layer4')
# l5_points(B,16,2048)
#################################################### c5
# c0_l5_xyz, c0_l5_points, c0_l5_indices = pointSIFT_res_module(l5_xyz, l5_points, radius=0.2, out_channel=2048,
# is_training=is_training, bn_decay=bn_decay,
# scope='layer5_c0',merge='concat')
# ###c0_15_points(B,64,2048)
#
# l6_xyz, l6_points, l6_indices = pointnet_sa_module(c0_l5_xyz, c0_l5_points, npoint=16, radius=0.2, nsample=32,
# mlp=[4096], mlp2=None, group_all=False,
# is_training=is_training, bn_decay=bn_decay, scope='layer5')
# l6_points(B,16,4096)
#################################################### c6
# c0_l6_xyz, c0_l6_points, c0_l6_indices = pointSIFT_res_module(l6_xyz, l6_points, radius=0.2, out_channel=4096,
# is_training=is_training, bn_decay=bn_decay,
# scope='layer6_c0', merge='concat')
# ###c0_16_points(B,16,4096)
#
# gp_xyz, gp_points, gp_indices = pointnet_sa_module(c0_l6_xyz, c0_l6_points, npoint=16, radius=0.2, nsample=32,
# mlp=[4096], mlp2=None, group_all=True,
# is_training=is_training, bn_decay=bn_decay, scope='layer6')
# # gp_points(B,1,4096)
#################################################### D E C O D E R #######################################################
# ugp_points=tf.expand_dims(gp_points, 1)#(B,1,1,4096)
#
# ugp_points=tf_util.conv2d_transpose(ugp_points,4096,[int(l6_points.shape[1]),1],padding='VALID',stride=[1,1],
# activation_fn=None, bn=True, is_training=is_training, scope='dconv1',
# bn_decay=bn_decay)
#
# ugp_points=tf.squeeze(ugp_points, [2])#(B,16,4096)
##################################################### stage 6
# u6_points_0 = PRRB(l6_xyz, l6_points, is_training, bn_decay,1)
#
# # u6_points(B,16,4096)
#
# c6_points = PCAB(u6_points_0, ugp_points,is_training, bn_decay,1)
#
# # c6_points(B,16,4096)
#
# u6_points_1 = PRRB(l6_xyz, c6_points, is_training, bn_decay, 2)
# u6_points(B,16,4096)
# l5_points_2 = pointnet_fp_module(l5_xyz, l6_xyz, l5_points, l6_points, [2048], is_training, bn_decay,
# scope='fp_layer6')
#
# ##################################################### stage 5
#
# u5_points_0 = PRRB(l5_xyz, l5_points, is_training, bn_decay, 3)
#
# # u5_points(B,64,2048)
#
# c5_points = PCAB(u5_points_0, l5_points_2, is_training, bn_decay, 2)
#
# # c5_points(B,64,2048)
#
# u5_points_1 = PRRB(l5_xyz, c5_points, is_training, bn_decay, 4)
#
# # u5_points(B,64,2048)
l4_points_2 = pointnet_fp_module(l4_xyz, l5_xyz, l4_points, l5_points, [1024], is_training, bn_decay,
scope='fp_layer5')
###################################################### stage 4
u4_points_0 = PRRB(l4_xyz, l4_points, is_training, bn_decay, 5)
# u4_points(B,128,1024)
c4_points = PCAB(u4_points_0, l4_points_2, is_training, bn_decay, 3)
# c4_points(B,128,1024)
u4_points_1 = PRRB(l4_xyz, c4_points, is_training, bn_decay, 6)
# u4_points(B,128,1024)
l3_points_2 = pointnet_fp_module(l3_xyz, l4_xyz, l3_points, u4_points_1, [512], is_training, bn_decay,
scope='fp_layer4')
####################################################### stage 3
u3_points_0 = PRRB(l3_xyz, l3_points, is_training, bn_decay, 7)
# u3_points(B,512,512)
c3_points = PCAB(u3_points_0, l3_points_2, is_training, bn_decay, 4)
# c3_points(B,512,512)
u3_points_1 = PRRB(l3_xyz, c3_points, is_training, bn_decay, 8)
# u3_points(B,512,512)
l2_points_2 = pointnet_fp_module(l2_xyz, l3_xyz, l2_points, u3_points_1, [256], is_training, bn_decay,
scope='fp_layer3')
######################################################## stage 2
u2_points_0 = PRRB(l2_xyz, l2_points, is_training, bn_decay, 9)
# u2_points(B,2048,256)
c2_points = PCAB(u2_points_0, l2_points_2, is_training, bn_decay, 5)
# c2_points(B,2048,256)
u2_points_1 = PRRB(l2_xyz, c2_points, is_training, bn_decay, 10)
# u2_points(B,2048,256)
l1_points_2 = pointnet_fp_module(l1_xyz, l2_xyz, l1_points, u2_points_1, [128], is_training, bn_decay,
scope='fp_layer2')
######################################################### stage 1
u1_points_0 = PRRB(l1_xyz, l1_points, is_training, bn_decay, 11)
# u1_points(B,8192,128)
c1_points = PCAB(u1_points_0, l1_points_2, is_training, bn_decay, 6)
# c1_points(B,8192,128)
u1_points_1 = PRRB(l1_xyz, c1_points, is_training, bn_decay, 12)
# u1_points(B,8192,128)
l0_points = pointnet_fp_module(l0_xyz, l1_xyz, l0_points, u1_points_1, [128, 128], is_training, bn_decay,
scope='fp_layer1')
_, l0_points, _ = pointSIFT_module(l0_xyz, l0_points, radius=0.1, out_channel=128, is_training=is_training,
bn_decay=bn_decay, scope='fsift_0')
######################################################### stage 0
# FC layers
net = tf_util.conv1d(l0_points, 128, 1, padding='VALID', bn=True, is_training=is_training, scope='fc1', bn_decay=bn_decay)
###l0_points(B,8192*5,128)
end_points['feats']=net
net = tf_util.dropout(net, keep_prob=0.5, is_training=is_training, scope='dp1')
net = tf_util.conv1d(net, num_class, 1, padding='VALID', activation_fn=None, scope='fc2')
extract_fea=tf.concat([end_points['l0_xyz'],end_points['feats']],axis=-1)
return net, extract_fea