github地址
PointNet 官方使用了 tensorflow
介绍
组成
1.PointNet classification network分类网络
- part segmentation network
数据集
1.point clouds sampled from 3D shapes
2.ShapeNetPart dataset.
结构
其主要分成以下三部分：
- 数据处理
- model构建
- 结果选择
数据处理
将点云处理成程序可用的格式，具体实现在 provider.py 中，主要包含了数据下载、预处理（shuffle->rotate->jitter）、格式转换（hdf5->txt）
shuffle
def shuffle_data(data, labels):
""" Shuffle data and labels.
Input:
data: B,N,... numpy array
label: B,... numpy array
Return:
shuffled data, label and shuffle indices
"""
idx = np.arange(len(labels))#返回一个列表
# print('idx=',idx)#idx= [ 0 1 2 ... 2045 2046 2047]
np.random.shuffle(idx)#把idx进行shuffle
# print('idx=', idx)
return data[idx, ...], labels[idx], idx
rotate旋转处理
def rotate_point_cloud(batch_data):
# print('batch data shape=',batch_data.shape)#(32, 1024, 3)
rotated_data = np.zeros(batch_data.shape, dtype=np.float32)
for k in range(batch_data.shape[0]):
rotation_angle = np.random.uniform() * 2 * np.pi#生成一个随机数
cosval = np.cos(rotation_angle)
sinval = np.sin(rotation_angle)
rotation_matrix = np.array([[cosval, 0, sinval],
[0, 1, 0],
[-sinval, 0, cosval]])
shape_pc = batch_data[k, ...]
rotated_data[k, ...] = np.dot(shape_pc.reshape((-1, 3)), rotation_matrix)
#先让shape_pc的形状变成(?,3),因为旋转矩阵为(3,3)
return rotated_data
解析：
>>>rotation_angle = np.random.uniform() * 2 * np.pi
>>> rotation_angle
3.1389077930878955
>>> cos=np.cos(rotation_angle)
>>> sin=np.sin(rotation_angle)
>>> cos
-0.9999963957642077
>>> sin
0.0026848572762735536
>>> matrix=np.array([[cos,0,sin],[0,1,0],[-sin,0,cos]]) >>> matrix
array([[-0.9999964 , 0. , 0.00268486],
[ 0. , 1. , 0. ],
[-0.00268486, 0. , -0.9999964 ]])
jitter抖动处理
def jitter_point_cloud(batch_data, sigma=0.01, clip=0.05):
B, N, C = batch_data.shape
assert(clip > 0)
jittered_data = np.clip(sigma * np.random.randn(B, N, C), -1*clip, clip)#将数组范围限制在(-1*clip, clip)
jittered_data += batch_data
return jittered_data
model构建
Feature transform net
with tf.variable_scope('transform_net1') as sc:#T-net
transform = input_transform_net(point_cloud, is_training, bn_decay, K=3)
print('point cloud=',point_cloud)#(32, 1024, 3)
# print('input transform=',transform)#(32, 3, 3)
point_cloud_transformed = tf.matmul(point_cloud, transform)
# print('point_cloud_transformed=',point_cloud_transformed)#(32, 1024, 3)
mlp(64,128,1024)
net = tf_util.conv2d(net_transformed, 64, [1,1],
padding='VALID', stride=[1,1],
bn=True, is_training=is_training,
scope='conv3', bn_decay=bn_decay)
print('net3=',net)#(32, 1024, 1, 64)
net = tf_util.conv2d(net, 128, [1,1],
padding='VALID', stride=[1,1],
bn=True, is_training=is_training,
scope='conv4', bn_decay=bn_decay)
print('net4=',net)#(32, 1024, 1, 128)
net = tf_util.conv2d(net, 1024, [1,1],
padding='VALID', stride=[1,1],
bn=True, is_training=is_training,
scope='conv5', bn_decay=bn_decay)
print('net5=',net)#(32, 1024, 1, 1024)
类别投票
实现方法
batch_pred_sum.shape=(?,40) # 每个data对40个类的可能性
pred_val.shape=(?,) # 每个data所属的可能性最大的类
pred_val = np.argmax(batch_pred_sum, 1)
#返回沿轴axis最大值的索引,即得到预测值最大的那一类的idx(label)
argmax()函数详解
例子１：
>>> a=np.arange(12).reshape(2,6)
>>> a
array([[ 0, 1, 2, 3, 4, 5],
[ 6, 7, 8, 9, 10, 11]])
>>> np.argmax(a,1)
array([5, 5]) #第五个数数值最大
例子２：
>>> a=np.arange(24).reshape(2,3,4)
>>> a
array([[[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]],
[[12, 13, 14, 15],
[16, 17, 18, 19],
[20, 21, 22, 23]]])
>>> np.argmax(a,1)
array([[2, 2, 2, 2],
[2, 2, 2, 2]])
>>> np.argmax(a,1).shape
(2, 4)
#[[ 0, 1, 2, 3],
# [ 4, 5, 6, 7],
# [ 8, 9, 10, 11]]中[ 8, 9, 10, 11]是最大的
>>> np.argmax(a,2)
array([[3, 3, 3], #[ 0, 1, 2, 3]中[3]是最大的
[3, 3, 3]])
>>> np.argmax(a,2).shape
(2, 3)
评估
输出（预测label，真实label）
</dump/pred_label.txt>
4, 4
0, 0
2, 2
8, 8
14, 23
...
<shape_names.txt>
airplane
bathtub
bed
bench
bookshelf
bottle
bowl
car
chair
cone
cup
保存预测错误的图片，并可视化
</dump/xxxx_pred_name.jpg>
命名＝第几个预测错误的图片+真实label+预测label
例子 /dump/1028_label_bed_pred_sofa.jpg
image.png
三张点云图片，分别是当前点云数据旋转三个不同角度之后的样子
save code
for i in range(start_idx, end_idx):
l = current_label[i]
total_seen_class[l] += 1
total_correct_class[l] += (pred_val[i-start_idx] == l)
fout.write('%d, %d\n' % (pred_val[i-start_idx], l))
# print('!!!!!!!!!!','%d, %d\n' % (pred_val[i-start_idx], l))
if pred_val[i-start_idx] != l and FLAGS.visu: # ERROR CASE, DUMP!如果预测错了
img_filename = '%d_label_%s_pred_%s.jpg' % (error_cnt, SHAPE_NAMES[l],
SHAPE_NAMES[pred_val[i-start_idx]])
#第几个预测错误的图片+真实label+预测label
img_filename = os.path.join(DUMP_DIR, img_filename)
output_img = pc_util.point_cloud_three_views(np.squeeze(current_data[i, :, :]))
scipy.misc.imsave(img_filename, output_img)
error_cnt += 1
画点云图的code
draw_point_cloud()
Input:
points: Nx3 numpy array
Output:
gray image
记录loss，预测精确度
/dump/log_evaluate.txt
eval mean loss: 1.816358
eval accuracy: 0.501216
eval avg class acc: 0.421297
airplane: 0.980
bathtub: 0.440
bed: 0.940
bench: 0.450
...
