def voting_module_simple(seed_xyz, seed_features):
'''
suppose vote_factor is 1
Input:
seed_xyz: (B, num_seed, 3) or (B, num_seed, 1, 3)
seed_features: (B, num_seed, seed_feature_dim) or (B, num_seed, 1, seed_feature_dim)
Returns:
vote_xyz: (B, num_seed, 1, 3)
vote_features: (B, num_seed, 1, seed_feature_dim)
'''
seed_feature_dim = K.int_shape(seed_features)[-1]
if len(K.int_shape(seed_xyz)) == 3:
seed_xyz = layers.Reshape((-1, 1, 3))(seed_xyz) #(B, num_seed, 1, 3)
seed_features = layers.Reshape((-1, 1, seed_feature_dim))(
seed_features) #(B, num_seed, 1, seed_feature_dim)
net = conv_bn_relu(seed_features, seed_feature_dim) # B, num_seed, 1, 3
net = conv_bn_relu(net, seed_feature_dim) # B, num_seed, 1, 3
net = layers.Conv2D(filters=(seed_feature_dim + 3), kernel_size=[1, 1])(
net) #(B, num_seed, 1, (3+seed_feature_dim))
# voting for the centroid of object
offset = crop(3, 0, 3)(net) # (B, num_seed, 1, 3)
vote_xyz = layers.Add()([seed_xyz, offset]) # (B, num_seed, 1, 3)
# feature residual
residual_features = crop(3, 3, seed_feature_dim + 3)(
net) # (B, num_seed, vote_factor, seed_feature_dim)
vote_features = layers.Add()(
[seed_features,
residual_features]) # # (B, num_seed, vote_factor, seed_feature_dim)
vote_features = layers.Reshape((-1, 1, seed_feature_dim))(vote_features)
return vote_xyz, vote_features
def voting_module(seed_xyz, seed_features, vote_factor):
'''
Inputs:
seed_xyz: (B, num_seed, 3) or (B, num_seed, 1, 3)
seed_features: (B, num_seed, seed_feature_dim) or (B, num_seed, 1, seed_feature_dim)
Returns:
vote_xyz: # (B, num_seed * vote_factor, 1, 3)
vote_features: (B, num_seed * vote_factor, 1, seed_feature_dim)
'''
seed_feature_dim = K.int_shape(seed_features)[-1]
if len(K.int_shape(seed_xyz)) == 3:
seed_xyz = layers.Reshape((-1, 1, 3))(seed_xyz) #(B, num_seed, 1, 3)
seed_features = layers.Reshape((-1, 1, seed_feature_dim))(
seed_features) #(B, num_seed, 1, seed_feature_dim)
num_seed = K.int_shape(seed_xyz)[-3]
num_vote = num_seed * vote_factor
net = conv_bn_relu(seed_features, seed_feature_dim) # B, num_seed, 1, 3
net = conv_bn_relu(net, seed_feature_dim) # B, num_seed, 1, 3
net = layers.Conv2D(filters=(seed_feature_dim + 3) * vote_factor,
kernel_size=[1, 1])(net)
# (B, num_seed, 1, (3+seed_feature_dim)*vote_factor)
if vote_factor != 1:
net = layers.Reshape((-1, vote_factor, 3 + seed_feature_dim))(
net) # (B, num_seed, vote_factor, (3+seed_feature_dim))
seed_xyz = layers.Lambda(K.tile,
arguments={'n': (1, 1, vote_factor, 1)})(
seed_xyz) # (B, num_seed, vote_factor, 3)
seed_features = layers.Lambda(
K.tile,
arguments={'n': (1, 1, vote_factor, 1)
})(seed_features) # (B,num_seed,vote_factor, dims)
# voting for the centroid of object
offset = crop(3, 0, 3)(net) # (B, num_seed, vote_factor, 3)
# vote_xyz = seed_xyz + offset # (B, num_seed, vote_factor, 3)
vote_xyz = layers.Add()([seed_xyz, offset])
vote_xyz = layers.Reshape(
(num_vote, 1, 3))(vote_xyz) # (B, num_seed * vote_factor, 1, 3)
# feature residual
residual_features = crop(3, 3, seed_feature_dim + 3)(
net) # (B, num_seed, vote_factor, seed_feature_dim)
# vote_features = seed_features + residual_features # # (B, num_seed, vote_factor, seed_feature_dim)
vote_features = layers.Add()([seed_features, residual_features])
vote_features = layers.Reshape((num_vote, 1, seed_feature_dim))(
vote_features) # (B, num_seed * vote_factor, 1, seed_feature_dim)
return vote_xyz, vote_features
def proposal_module(vote_xyz,
vote_features,
num_class,
num_head_bin,
num_size_cluster,
num_proposal,
random_sample=False):
'''
Return:
xyz: proposal_xyz
scores: scores
idx: proposal index
'''
seed_feature_dim = K.int_shape(vote_features)[-1]
vote_xyz = layers.Reshape((-1, 3))(vote_xyz)
vote_features = layers.Reshape(
(-1, seed_feature_dim))(vote_features) # squeeze the tensors
xyz, features, idx = pointnet_sa_module(vote_xyz,
vote_features,
mlp=[128, 128, 128],
n_centroid=num_proposal,
n_samples=16,
radius=0.3,
use_feature=True,
use_xyz=True,
random_sample=random_sample)
features = layers.Reshape((-1, 1, 128))(features) # 128 = mlp[-1]
net = conv_bn_relu(features, 128)
net = conv_bn_relu(net, 128)
# objectness scores: 2
# center residual: 3
# heading: classification + residual = 2*num_head_bin
# size: classification + residual = 4*num_size_cluster
# object classification: num_class
output_dims = 2 + 3 + num_head_bin * 2 + num_size_cluster * 4 + num_class
scores = layers.Conv2D(filters=output_dims, kernel_size=(1, 1))(net)
return xyz, scores, idx