-
Notifications
You must be signed in to change notification settings - Fork 361
/
pointnet.py
284 lines (246 loc) · 9.57 KB
/
pointnet.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
# Copyright (c) Chris Choy (chrischoy@ai.stanford.edu).
#
# Permission is hereby granted, free of charge, to any person obtaining a copy of
# this software and associated documentation files (the "Software"), to deal in
# the Software without restriction, including without limitation the rights to
# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
# of the Software, and to permit persons to whom the Software is furnished to do
# so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
#
# Please cite "4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural
# Networks", CVPR'19 (https://arxiv.org/abs/1904.08755) if you use any part
# of the code.
import os
import numpy as np
from urllib.request import urlretrieve
try:
import open3d as o3d
except ImportError:
raise ImportError('Please install open3d with `pip install open3d`.')
import torch
import torch.nn as nn
import MinkowskiEngine as ME
class STN3d(nn.Module):
r"""Given a sparse tensor, generate a 3x3 transformation matrix per
instance.
"""
CONV_CHANNELS = [64, 128, 1024, 512, 256]
FC_CHANNELS = [512, 256]
KERNEL_SIZES = [1, 1, 1]
STRIDES = [1, 1, 1]
def __init__(self, D=3):
super(STN3d, self).__init__()
k = self.KERNEL_SIZES
s = self.STRIDES
c = self.CONV_CHANNELS
self.block1 = nn.Sequential(
ME.MinkowskiConvolution(
3,
c[0],
kernel_size=k[0],
stride=s[0],
has_bias=False,
dimension=3), ME.MinkowskiInstanceNorm(c[0]),
ME.MinkowskiReLU())
self.block2 = nn.Sequential(
ME.MinkowskiConvolution(
c[0],
c[1],
kernel_size=k[1],
stride=s[1],
has_bias=False,
dimension=3), ME.MinkowskiInstanceNorm(c[1]),
ME.MinkowskiReLU())
self.block3 = nn.Sequential(
ME.MinkowskiConvolution(
c[1],
c[2],
kernel_size=k[2],
stride=s[2],
has_bias=False,
dimension=3), ME.MinkowskiInstanceNorm(c[2]),
ME.MinkowskiReLU())
# Use the kernelsize 1 convolution for linear layers. If kernel size ==
# 1, minkowski engine internally uses a linear function.
self.block4 = nn.Sequential(
ME.MinkowskiConvolution(
c[2], c[3], kernel_size=1, has_bias=False, dimension=3),
ME.MinkowskiInstanceNorm(c[3]), ME.MinkowskiReLU())
self.block5 = nn.Sequential(
ME.MinkowskiConvolution(
c[3], c[4], kernel_size=1, has_bias=False, dimension=3),
ME.MinkowskiInstanceNorm(c[4]), ME.MinkowskiReLU())
self.fc6 = ME.MinkowskiConvolution(
c[4], 9, kernel_size=1, has_bias=True, dimension=3)
self.avgpool = ME.MinkowskiGlobalPooling()
self.broadcast = ME.MinkowskiBroadcast()
def forward(self, in_x):
x = self.block1(in_x)
x = self.block2(x)
x = self.block3(x)
# batch size x channel
x = self.avgpool(x)
x = self.block4(x)
x = self.block5(x)
# get the features batch-wise
x = self.fc6(x)
# Add identity transformation
x._F += torch.tensor([[1, 0, 0, 0, 1, 0, 0, 0, 1]],
dtype=x.dtype,
device=x.device).repeat(len(x), 1)
# Broadcast the transformation back to the right coordinates of x
return self.broadcast(in_x, x)
class PointNetFeature(nn.Module):
r"""
You can think of a PointNet as a specialization of a convolutional neural
network with kernel_size == 1, and stride == 1 that processes a sparse
tensor where features are normalized coordinates.
This generalization allows the network to process an arbitrary number of
points.
"""
CONV_CHANNELS = [256, 512, 1024]
KERNEL_SIZES = [1, 1, 1]
STRIDES = [1, 1, 1]
def __init__(self):
super(PointNetFeature, self).__init__()
k = self.KERNEL_SIZES
s = self.STRIDES
c = self.CONV_CHANNELS
self.stn = STN3d(D=3)
self.block1 = nn.Sequential(
ME.MinkowskiConvolution(
6,
c[0],
kernel_size=k[0],
stride=s[0],
has_bias=False,
dimension=3), ME.MinkowskiInstanceNorm(c[0]),
ME.MinkowskiReLU())
self.block2 = nn.Sequential(
ME.MinkowskiConvolution(
c[0],
c[1],
kernel_size=k[1],
stride=s[1],
has_bias=False,
dimension=3), ME.MinkowskiInstanceNorm(c[1]),
ME.MinkowskiReLU())
self.block3 = nn.Sequential(
ME.MinkowskiConvolution(
c[1],
c[2],
kernel_size=k[2],
stride=s[2],
has_bias=False,
dimension=3), ME.MinkowskiInstanceNorm(c[2]),
ME.MinkowskiReLU())
self.avgpool = ME.MinkowskiGlobalPooling()
self.concat = ME.MinkowskiBroadcastConcatenation()
def forward(self, x):
"""
Input is a spare tensor with features as centered coordinates N x 3
"""
assert isinstance(x, ME.SparseTensor)
assert x.F.shape[1] == 3
# Get the transformation
T = self.stn(x)
# Apply the transformation
coords_feat_stn = torch.squeeze(
torch.bmm(x.F.view(-1, 1, 3), T.F.view(-1, 3, 3)))
x = ME.SparseTensor(
torch.cat((coords_feat_stn, x.F), 1),
coords_key=x.coords_key,
coords_manager=x.coords_man)
point_feat = self.block1(x)
x = self.block2(point_feat)
x = self.block3(x)
glob_feat = self.avgpool(x)
return self.concat(point_feat, glob_feat)
class PointNet(nn.Module):
r"""
You can think of a PointNet as a specialization of a convolutional neural
network with kernel_size == 1, and stride == 1 that processes a sparse
tensor where features are normalized coordinates.
This generalization allows the network to process an arbitrary number of
points.
"""
CONV_CHANNELS = [512, 256, 128]
KERNEL_SIZES = [1, 1, 1]
STRIDES = [1, 1, 1]
def __init__(self, out_channels, D=3):
super(PointNet, self).__init__()
k = self.KERNEL_SIZES
s = self.STRIDES
c = self.CONV_CHANNELS
self.feat = PointNetFeature()
self.block1 = nn.Sequential(
ME.MinkowskiConvolution(
1280,
c[0],
kernel_size=k[0],
stride=s[0],
has_bias=False,
dimension=3), ME.MinkowskiInstanceNorm(c[0]),
ME.MinkowskiReLU())
self.block2 = nn.Sequential(
ME.MinkowskiConvolution(
c[0],
c[1],
kernel_size=k[1],
stride=s[1],
has_bias=False,
dimension=3), ME.MinkowskiInstanceNorm(c[1]),
ME.MinkowskiReLU())
self.block3 = nn.Sequential(
ME.MinkowskiConvolution(
c[1],
c[2],
kernel_size=k[2],
stride=s[2],
has_bias=False,
dimension=3), ME.MinkowskiInstanceNorm(c[2]),
ME.MinkowskiReLU())
# Last FC layer. Note that kernel_size 1 == linear layer
self.conv4 = ME.MinkowskiConvolution(
c[2], out_channels, kernel_size=1, has_bias=True, dimension=3)
def forward(self, x):
"""
Assume that x.F (features) are normalized coordinates or centered coordinates
"""
assert isinstance(x, ME.SparseTensor)
x = self.feat(x)
x = self.block1(x)
x = self.block2(x)
x = self.block3(x)
return self.conv4(x)
bunny_file = "bunny.ply"
if not os.path.isfile(bunny_file):
urlretrieve(
"https://raw.githubusercontent.com/naucoin/VTKData/master/Data/bunny.ply",
bunny_file)
if __name__ == '__main__':
voxel_size = 2e-3 # High resolution grid works better just like high-res image is better for 2D classification
pointnet = PointNet(20).float()
pcd = o3d.io.read_point_cloud(bunny_file)
# If you need a high-resolution point cloud, sample points using
# https://chrischoy.github.io/research/barycentric-coordinate-for-mesh-sampling/
# Convert to a voxel grid
coords = np.array(pcd.points)
feats = coords - coords.mean(0) # Coordinates are features for pointnet
quantized_coords = np.floor(coords / voxel_size)
inds = ME.utils.sparse_quantize(quantized_coords, return_index=True)
quantized_coords, feats = ME.utils.sparse_collate([quantized_coords[inds]],
[feats[inds]])
sinput = ME.SparseTensor(feats.float(), quantized_coords)
print(pointnet(sinput))