-
Notifications
You must be signed in to change notification settings - Fork 1
/
kit.py
355 lines (289 loc) · 11.9 KB
/
kit.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
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
import os
import math
import multiprocessing
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import torch
import torch.nn as nn
from tqdm import tqdm
from plyfile import PlyElement, PlyData
from pyntcloud import PyntCloud
from pytorch3d.ops.knn import knn_gather, knn_points
#core transformation function
def transformRGBToYCoCg(bitdepth, rgb):
r = rgb[:, 0]
g = rgb[:, 1]
b = rgb[:, 2]
co = r - b
t = b + (co >> 1) # co >>1 i.e. co // 2
cg = g - t
y = t + (cg >> 1)
offset = 1 << bitdepth # 2^bitdepth
# NB: YCgCoR needs extra 1-bit for chroma
return np.column_stack((y, co + offset, cg + offset))
def transformYCoCgToRGB(bitDepth, ycocg):
offset = 1 << bitDepth
y0 = ycocg[:,0]
co = ycocg[:,1] - offset
cg = ycocg[:,2] - offset
t = y0 - (cg >> 1)
g = cg + t
b = t - (co >> 1)
r = co + b
maxVal = (1 << bitDepth) - 1
r = np.clip(r, 0, maxVal)
g = np.clip(g, 0, maxVal)
b = np.clip(b, 0, maxVal)
return np.column_stack((r,g,b))
def read_point_cloud_ycocg(filepath):
pc = PyntCloud.from_file(filepath)
try:
cols=['x', 'y', 'z','red', 'green', 'blue']
points=pc.points[cols].values
except:
cols = ['x', 'y', 'z', 'r', 'g', 'b']
points = pc.points[cols].values
color = points[:, 3:].astype(np.int16)
color = transformRGBToYCoCg(8, color)
# color: int
# y channel: 0~255
# co channel: 0~511 (1~511 in our dataset)
# cg channel: 0~511 (34~476 in our dataset)
points[:, 3:] = color.astype(float)
return points
def save_point_cloud_ycocg(pc, path):
color = pc[:, 3:]
color = np.round(color).astype(np.int16) # 务必 round 后 再加 astype
color = transformYCoCgToRGB(8, color)
pc = pd.DataFrame(pc, columns=['x', 'y', 'z', 'red', 'green', 'blue'])
pc[['red','green','blue']] = np.round(color).astype(np.uint8)
cloud = PyntCloud(pc)
cloud.to_file(path)
def read_point_cloud_reflactance(filepath):
plydata = PlyData.read(filepath)
pc = np.array(np.transpose(np.stack((plydata['vertex']['x'],plydata['vertex']['y'],plydata['vertex']['z'], plydata['vertex']['reflectance'])))).astype(np.float32)
return pc
def save_point_cloud_reflactance(pc, path, to_rgb=False):
if to_rgb:
pc[:, 3:] = pc[:, 3:] / 100
cmap = plt.get_cmap('jet')
color = np.round(cmap(pc[:, 3])[:, :3] * 255)
pc = np.hstack((pc[:, :3], color))
pc = pd.DataFrame(pc, columns=['x', 'y', 'z', 'red', 'green', 'blue'])
pc[['red','green','blue']] = np.round(np.clip(pc[['red','green','blue']], 0, 255)).astype(np.uint8)
cloud = PyntCloud(pc)
cloud.to_file(path)
else:
scan = pc
vertex = np.array(
[(scan[i,0], scan[i,1], scan[i,2], scan[i,3]) for i in range(scan.shape[0])],
dtype=[
("x", np.dtype("float32")),
("y", np.dtype("float32")),
("z", np.dtype("float32")),
("reflectance", np.dtype("uint8")),
]
)
PlyElement.describe(vertex, 'vertex', comments=['vertices'])
output_pc = PlyElement.describe(vertex, "vertex")
output_pc = PlyData([output_pc])
output_pc.write(path)
def read_point_clouds_ycocg(file_path_list, bar=True):
print('loading point clouds...')
with multiprocessing.Pool() as p:
if bar:
pcs = list(tqdm(p.imap(read_point_cloud_ycocg, file_path_list, 32), total=len(file_path_list)))
else:
pcs = list(p.imap(read_point_cloud_ycocg, file_path_list, 32))
return pcs
def n_scale_ball(grouped_xyz):
B, N, K, _ = grouped_xyz.shape
longest = (grouped_xyz**2).sum(dim=-1).sqrt().max(dim=-1)[0]
scaling = (1) / longest
grouped_xyz = grouped_xyz * scaling.view(B, N, 1, 1)
return grouped_xyz
class MLP(nn.Module):
def __init__(self, in_channel, mlp, relu, bn):
super(MLP, self).__init__()
mlp.insert(0, in_channel)
self.mlp_Modules = nn.ModuleList()
for i in range(len(mlp) - 1):
if relu[i]:
if bn[i]:
mlp_Module = nn.Sequential(
nn.Conv2d(mlp[i], mlp[i+1], 1),
nn.BatchNorm2d(mlp[i+1]),
nn.ReLU(),
)
else:
mlp_Module = nn.Sequential(
nn.Conv2d(mlp[i], mlp[i+1], 1),
nn.ReLU(),
)
else:
mlp_Module = nn.Sequential(
nn.Conv2d(mlp[i], mlp[i+1], 1),
)
self.mlp_Modules.append(mlp_Module)
def forward(self, points, squeeze=False):
"""
Input:
points: input points position data, [B, C, N]
Return:
points: feature data, [B, D, N]
"""
if squeeze:
points = points.unsqueeze(-1) # [B, C, N, 1]
for m in self.mlp_Modules:
points = m(points)
# [B, D, N, 1]
if squeeze:
points = points.squeeze(-1) # [B, D, N]
return points
class QueryMaskedAttention(nn.Module):
def __init__(self, channel):
super(QueryMaskedAttention, self).__init__()
self.channel = channel
self.k_mlp = nn.Conv2d(in_channels=channel, out_channels=channel, kernel_size=1)
self.v_mlp = nn.Conv2d(in_channels=channel, out_channels=channel, kernel_size=1)
self.pe_multiplier, self.pe_bias = True, True
if self.pe_multiplier:
self.linear_p_multiplier = nn.Sequential(
nn.Conv2d(in_channels=3, out_channels=channel, kernel_size=1),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=channel, out_channels=channel, kernel_size=1),
)
if self.pe_bias:
self.linear_p_bias = nn.Sequential(
nn.Conv2d(in_channels=3, out_channels=channel, kernel_size=1),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=channel, out_channels=channel, kernel_size=1),
)
self.weight_encoding = nn.Sequential(
nn.Conv2d(in_channels=channel, out_channels=channel, kernel_size=1),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=channel, out_channels=channel, kernel_size=1),
)
self.residual_emb = nn.Sequential(
nn.ReLU(),
nn.Conv2d(in_channels=channel, out_channels=channel, kernel_size=1),
)
self.softmax = nn.Softmax(dim=2)
def forward(self, grouped_xyz, grouped_feature):
key = self.k_mlp(grouped_feature) # B, C, K, M
value = self.v_mlp(grouped_feature) # B, C, K, M
relation_qk = key # - query
if self.pe_multiplier:
pem = self.linear_p_multiplier(grouped_xyz)
relation_qk = relation_qk * pem
if self.pe_bias:
peb = self.linear_p_bias(grouped_xyz)
relation_qk = relation_qk + peb
value = value + peb
weight = self.weight_encoding(relation_qk)
score = self.softmax(weight) # B, C, K, M
feature = score*value # B, C, K, M
feature = self.residual_emb(feature) # B, C, K, M
return feature
class PT(nn.Module):
def __init__(self, in_channel, out_channel, n_layers):
super(PT, self).__init__()
self.in_channel = in_channel
self.out_channel = out_channel
self.n_layers = n_layers
self.sa_ls, self.sa_emb_ls = nn.ModuleList(), nn.ModuleList()
self.linear_in = nn.Conv2d(in_channel, out_channel, kernel_size=1)
for i in range(n_layers):
self.sa_emb_ls.append(nn.Sequential(
nn.Conv2d(out_channel, out_channel, kernel_size=1),
nn.ReLU(),
))
self.sa_ls.append(QueryMaskedAttention(out_channel))
def forward(self, groped_geo, grouped_attr):
"""
Input:
groped_geo: input points position data, [B, M, K, 3]
groped_attr: input points feature data, [B, M, K, 3]
Return:
feature: output feature data, [B, M, C]
"""
groped_geo, grouped_attr = groped_geo.permute((0, 3, 2, 1)), grouped_attr.permute((0, 3, 2, 1)) # B, _, K, M
feature = self.linear_in(grouped_attr)
for i in range(self.n_layers):
identity = feature
feature = self.sa_emb_ls[i](feature)
output = self.sa_ls[i](groped_geo, feature)
feature = output + identity
feature = feature.sum(dim=2).transpose(1, 2)
return feature
def get_cdf(mu, sigma):
M, d = sigma.shape
mu = mu.unsqueeze(-1).repeat(1, 1, 256)
sigma = sigma.unsqueeze(-1).repeat(1, 1, 256).clamp(1e-10, 1e10)
gaussian = torch.distributions.laplace.Laplace(mu, sigma)
flag = torch.arange(0, 256).to(sigma.device).view(1, 1, 256).repeat((M, d, 1))
cdf = gaussian.cdf(flag + 0.5)
spatial_dimensions = cdf.shape[:-1] + (1,)
zeros = torch.zeros(spatial_dimensions, dtype=cdf.dtype, device=cdf.device)
cdf_with_0 = torch.cat([zeros, cdf], dim=-1)
return cdf_with_0
def get_cdf_ycocg(mu, sigma):
M, d = sigma.shape
mu = mu.unsqueeze(-1).repeat(1, 1, 512)
sigma = sigma.unsqueeze(-1).repeat(1, 1, 512).clamp(1e-10, 1e10)
gaussian = torch.distributions.laplace.Laplace(mu, sigma)
flag = torch.arange(0, 512).to(sigma.device).view(1, 1, 512).repeat((M, d, 1))
cdf = gaussian.cdf(flag + 0.5)
spatial_dimensions = cdf.shape[:-1] + (1,)
zeros = torch.zeros(spatial_dimensions, dtype=cdf.dtype, device=cdf.device)
cdf_with_0 = torch.cat([zeros, cdf], dim=-1)
return cdf_with_0
def get_cdf_reflactance(mu, sigma):
M, d = sigma.shape
mu = mu.unsqueeze(-1).repeat(1, 1, 128)
sigma = sigma.unsqueeze(-1).repeat(1, 1, 128).clamp(1e-10, 1e10)
gaussian = torch.distributions.laplace.Laplace(mu, sigma)
flag = torch.arange(0, 128).to(sigma.device).view(1, 1, 128).repeat((M, d, 1))
cdf = gaussian.cdf(flag + 0.5)
spatial_dimensions = cdf.shape[:-1] + (1,)
zeros = torch.zeros(spatial_dimensions, dtype=cdf.dtype, device=cdf.device)
cdf_with_0 = torch.cat([zeros, cdf], dim=-1)
return cdf_with_0
def feature_probs_based_mu_sigma(feature, mu, sigma):
sigma = sigma.clamp(1e-10, 1e10)
gaussian = torch.distributions.laplace.Laplace(mu, sigma)
probs = gaussian.cdf(feature + 0.5) - gaussian.cdf(feature - 0.5)
total_bits = torch.sum(torch.clamp(-1.0 * torch.log(probs + 1e-10) / math.log(2.0), 0, 50))
return total_bits, probs
def get_file_size_in_bits(f):
return os.stat(f).st_size * 8
def _convert_to_int_and_normalize(cdf_float, needs_normalization):
"""Convert floatingpoint CDF to integers. See README for more info.
The idea is the following:
When we get the cdf here, it is (assumed to be) between 0 and 1, i.e,
cdf \in [0, 1)
(note that 1 should not be included.)
We now want to convert this to int16 but make sure we do not get
the same value twice, as this would break the arithmetic coder
(you need a strictly monotonically increasing function).
So, if needs_normalization==True, we multiply the input CDF
with 2**16 - (Lp - 1). This means that now,
cdf \in [0, 2**16 - (Lp - 1)].
Then, in a final step, we add an arange(Lp), which is just a line with
slope one. This ensure that for sure, we will get unique, strictly
monotonically increasing CDFs, which are \in [0, 2**16)
"""
Lp = cdf_float.shape[-1]
factor = torch.tensor(
2, dtype=torch.float32, device=cdf_float.device).pow_(16)
new_max_value = factor
if needs_normalization:
new_max_value = new_max_value - (Lp - 1)
cdf_float = cdf_float.mul(new_max_value)
cdf_float = cdf_float.round()
cdf = cdf_float.to(dtype=torch.int16, non_blocking=True)
if needs_normalization:
r = torch.arange(Lp, dtype=torch.int16, device=cdf.device)
cdf.add_(r)
return cdf