def pointSIFT_KNN(radius, xyz):
idx = pointSIFT_select_four(xyz, radius)
grouped_xyz = group_point(xyz, idx) # (batch_size, npoint, 32, 3)
xyz_central = tf.tile(tf.expand_dims(xyz, 2),
[1, 1, 32, 1]) # translation normalization
a = grouped_xyz - xyz_central
new_points = tf.concat([xyz_central, a], axis=-1)
return new_points, idx
def pointSIFT_group_four(radius, xyz, points, use_xyz=True):
idx = pointSIFT_select_four(xyz, radius)
grouped_xyz = group_point(xyz, idx) # (batch_size, npoint, 32, 3)
grouped_xyz -= tf.tile(tf.expand_dims(xyz, 2), [1, 1, 32, 1]) # translation normalization
if points is not None:
grouped_points = group_point(points, idx) # (batch_size, npoint, 8/32, channel)
if use_xyz:
new_points = tf.concat([grouped_xyz, grouped_points], axis=-1) # (batch_size, npoint, 8/32, 3+channel)
else:
new_points = grouped_points
else:
new_points = grouped_xyz
return xyz, new_points, idx, grouped_xyz
def get_model(point_input, is_training, pfs_flag=False, bn_decay=None):
""" Classification PointNet, input is BxNxC, output Bx40 """
batch_size = point_input.get_shape()[0].value
num_point1 = point_input.get_shape()[1].value
num_point2 = int(np.floor(num_point1 / 4.0))
num_point3 = int(np.floor(num_point2 / 4.0))
num_features = point_input.get_shape()[2].value
end_points = {}
point_cloud1 = point_input
k = 32
nn_idx = pointSIFT_select_four(point_cloud1, 0.2)
net1_1 = tf_util.attention_conv(point_cloud1,
point_input,
64,
nn_idx,
k,
scope='conv_1_1',
bn=True,
bn_decay=bn_decay,
is_training=is_training)
net1_2 = tf_util.attention_conv(point_cloud1,
net1_1,
64,
nn_idx,
k,
scope='conv_1_2',
bn=True,
bn_decay=bn_decay,
is_training=is_training)
k = 30
net, p1_idx, pn_idx, point_cloud2 = tf_util.attention_pooling(
net1_2,
point_cloud1,
num_point2,
k,
scope='12',
bn_decay=bn_decay,
is_training=is_training)
net1_1 = tf.squeeze(
tf.reduce_max(group_point(net1_1, pn_idx), axis=-2, keepdims=True))
net1_2 = net
k = 16
nn_idx = pointSIFT_select_two(point_cloud2, 0.4)
net2_1 = tf_util.attention_conv(point_cloud2,
net,
128,
nn_idx,
k,
scope='conv_2_1',
bn=True,
bn_decay=bn_decay,
is_training=is_training)
net2_2 = tf_util.attention_conv(point_cloud2,
net2_1,
128,
nn_idx,
k,
scope='conv_2_2',
bn=True,
bn_decay=bn_decay,
is_training=is_training)
k = 30
net, p2_idx, pn_idx, point_cloud3 = tf_util.attention_pooling(
net2_2,
point_cloud2,
num_point3,
k,
scope='13',
bn_decay=bn_decay,
is_training=is_training)
print(6666, net1_1.shape)
net1_1 = tf.reduce_max(group_point(net1_1, pn_idx), axis=-2, keepdims=True)
net1_2 = tf.reduce_max(group_point(net1_2, pn_idx), axis=-2, keepdims=True)
net2_1 = tf.reduce_max(group_point(net2_1, pn_idx), axis=-2, keepdims=True)
net2_2 = net
k = 16
nn_idx = pointSIFT_select_two(point_cloud3, 0.6)
net3_1 = tf_util.attention_conv(point_cloud3,
net,
256,
nn_idx,
k,
scope='conv_3_1',
bn=True,
bn_decay=bn_decay,
is_training=is_training)
net3_2 = tf_util.attention_conv(point_cloud3,
net3_1,
256,
nn_idx,
k,
scope='conv_3_2',
bn=True,
bn_decay=bn_decay,
is_training=is_training)
net3_1 = tf.expand_dims(net3_1, axis=-2)
net3_2 = tf.expand_dims(net3_2, axis=-2)
net2_2 = tf.expand_dims(net2_2, axis=-2)
net = tf.concat([net1_1, net1_2, net2_1, net2_2, net3_1, net3_2], axis=-1)
net = tf_util.conv2d(net,
1024, [1, 1],
padding='VALID',
stride=[1, 1],
activation_fn=tf.nn.relu,
bn=True,
is_training=is_training,
scope='agg',
bn_decay=bn_decay)
net = tf.reduce_max(net, axis=1, keepdims=True)
net = tf.reshape(net, [batch_size, -1])
end_points['embedding'] = net
net = tf_util.fully_connected(net,
512,
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='dropout1')
net = tf_util.fully_connected(net,
256,
bn=True,
is_training=is_training,
scope='fc2',
bn_decay=bn_decay)
net = tf_util.dropout(net,
keep_prob=0.5,
is_training=is_training,
scope='dropout2')
net = tf_util.fully_connected(net, 40, activation_fn=None, scope='fc3')
return net, end_points