Update pages/01_🦷 Segment.py
Browse files- pages/01_🦷 Segment.py +131 -157
pages/01_🦷 Segment.py
CHANGED
@@ -1,27 +1,29 @@
|
|
1 |
-
|
2 |
import shutil
|
|
|
3 |
|
4 |
-
import os
|
5 |
import numpy as np
|
6 |
-
from sklearn import neighbors
|
7 |
from scipy.spatial import distance_matrix
|
|
|
8 |
from pygco import cut_from_graph
|
9 |
import open3d as o3d
|
10 |
import matplotlib.pyplot as plt
|
11 |
import matplotlib.colors as mcolors
|
12 |
-
from stqdm import stqdm
|
13 |
-
import json
|
14 |
-
from stpyvista import stpyvista
|
15 |
import torch
|
16 |
import torch.nn as nn
|
17 |
from torch.autograd import Variable
|
18 |
import torch.nn.functional as F
|
19 |
import streamlit as st
|
20 |
-
import
|
21 |
-
|
|
|
22 |
from PIL import Image
|
23 |
|
|
|
24 |
class TeethApp:
|
|
|
|
|
|
|
25 |
def __init__(self):
|
26 |
# Font
|
27 |
with open("utils/style.css") as css:
|
@@ -48,44 +50,6 @@ class TeethApp:
|
|
48 |
unsafe_allow_html=True,
|
49 |
)
|
50 |
|
51 |
-
|
52 |
-
class STN3d(nn.Module):
|
53 |
-
def __init__(self, channel):
|
54 |
-
super(STN3d, self).__init__()
|
55 |
-
self.conv1 = torch.nn.Conv1d(channel, 64, 1)
|
56 |
-
self.conv2 = torch.nn.Conv1d(64, 128, 1)
|
57 |
-
self.conv3 = torch.nn.Conv1d(128, 1024, 1)
|
58 |
-
self.fc1 = nn.Linear(1024, 512)
|
59 |
-
self.fc2 = nn.Linear(512, 256)
|
60 |
-
self.fc3 = nn.Linear(256, 9)
|
61 |
-
self.relu = nn.ReLU()
|
62 |
-
|
63 |
-
self.bn1 = nn.BatchNorm1d(64)
|
64 |
-
self.bn2 = nn.BatchNorm1d(128)
|
65 |
-
self.bn3 = nn.BatchNorm1d(1024)
|
66 |
-
self.bn4 = nn.BatchNorm1d(512)
|
67 |
-
self.bn5 = nn.BatchNorm1d(256)
|
68 |
-
|
69 |
-
def forward(self, x):
|
70 |
-
batchsize = x.size()[0]
|
71 |
-
x = F.relu(self.bn1(self.conv1(x)))
|
72 |
-
x = F.relu(self.bn2(self.conv2(x)))
|
73 |
-
x = F.relu(self.bn3(self.conv3(x)))
|
74 |
-
x = torch.max(x, 2, keepdim=True)[0]
|
75 |
-
x = x.view(-1, 1024)
|
76 |
-
|
77 |
-
x = F.relu(self.bn4(self.fc1(x)))
|
78 |
-
x = F.relu(self.bn5(self.fc2(x)))
|
79 |
-
x = self.fc3(x)
|
80 |
-
|
81 |
-
iden = Variable(torch.from_numpy(np.array([1, 0, 0, 0, 1, 0, 0, 0, 1]).astype(np.float32))).view(1, 9).repeat(
|
82 |
-
batchsize, 1)
|
83 |
-
if x.is_cuda:
|
84 |
-
iden = iden.to(x.get_device())
|
85 |
-
x = x + iden
|
86 |
-
x = x.view(-1, 3, 3)
|
87 |
-
return x
|
88 |
-
|
89 |
class STNkd(nn.Module):
|
90 |
def __init__(self, k=64):
|
91 |
super(STNkd, self).__init__()
|
@@ -133,13 +97,15 @@ class MeshSegNet(nn.Module):
|
|
133 |
self.with_dropout = with_dropout
|
134 |
self.dropout_p = dropout_p
|
135 |
|
136 |
-
# MLP-1 [64, 64]
|
137 |
self.mlp1_conv1 = torch.nn.Conv1d(self.num_channels, 64, 1)
|
138 |
self.mlp1_conv2 = torch.nn.Conv1d(64, 64, 1)
|
139 |
self.mlp1_bn1 = nn.BatchNorm1d(64)
|
140 |
self.mlp1_bn2 = nn.BatchNorm1d(64)
|
|
|
141 |
# FTM (feature-transformer module)
|
142 |
self.fstn = STNkd(k=64)
|
|
|
143 |
# GLM-1 (graph-contrained learning modulus)
|
144 |
self.glm1_conv1_1 = torch.nn.Conv1d(64, 32, 1)
|
145 |
self.glm1_conv1_2 = torch.nn.Conv1d(64, 32, 1)
|
@@ -147,6 +113,7 @@ class MeshSegNet(nn.Module):
|
|
147 |
self.glm1_bn1_2 = nn.BatchNorm1d(32)
|
148 |
self.glm1_conv2 = torch.nn.Conv1d(32+32, 64, 1)
|
149 |
self.glm1_bn2 = nn.BatchNorm1d(64)
|
|
|
150 |
# MLP-2
|
151 |
self.mlp2_conv1 = torch.nn.Conv1d(64, 64, 1)
|
152 |
self.mlp2_bn1 = nn.BatchNorm1d(64)
|
@@ -154,6 +121,7 @@ class MeshSegNet(nn.Module):
|
|
154 |
self.mlp2_bn2 = nn.BatchNorm1d(128)
|
155 |
self.mlp2_conv3 = torch.nn.Conv1d(128, 512, 1)
|
156 |
self.mlp2_bn3 = nn.BatchNorm1d(512)
|
|
|
157 |
# GLM-2 (graph-contrained learning modulus)
|
158 |
self.glm2_conv1_1 = torch.nn.Conv1d(512, 128, 1)
|
159 |
self.glm2_conv1_2 = torch.nn.Conv1d(512, 128, 1)
|
@@ -163,6 +131,7 @@ class MeshSegNet(nn.Module):
|
|
163 |
self.glm2_bn1_3 = nn.BatchNorm1d(128)
|
164 |
self.glm2_conv2 = torch.nn.Conv1d(128*3, 512, 1)
|
165 |
self.glm2_bn2 = nn.BatchNorm1d(512)
|
|
|
166 |
# MLP-3
|
167 |
self.mlp3_conv1 = torch.nn.Conv1d(64+512+512+512, 256, 1)
|
168 |
self.mlp3_conv2 = torch.nn.Conv1d(256, 256, 1)
|
@@ -172,7 +141,8 @@ class MeshSegNet(nn.Module):
|
|
172 |
self.mlp3_conv4 = torch.nn.Conv1d(128, 128, 1)
|
173 |
self.mlp3_bn2_1 = nn.BatchNorm1d(128)
|
174 |
self.mlp3_bn2_2 = nn.BatchNorm1d(128)
|
175 |
-
|
|
|
176 |
self.output_conv = torch.nn.Conv1d(128, self.num_classes, 1)
|
177 |
if self.with_dropout:
|
178 |
self.dropout = nn.Dropout(p=self.dropout_p)
|
@@ -180,13 +150,16 @@ class MeshSegNet(nn.Module):
|
|
180 |
def forward(self, x, a_s, a_l):
|
181 |
batchsize = x.size()[0]
|
182 |
n_pts = x.size()[2]
|
|
|
183 |
# MLP-1
|
184 |
x = F.relu(self.mlp1_bn1(self.mlp1_conv1(x)))
|
185 |
x = F.relu(self.mlp1_bn2(self.mlp1_conv2(x)))
|
|
|
186 |
# FTM
|
187 |
trans_feat = self.fstn(x)
|
188 |
x = x.transpose(2, 1)
|
189 |
x_ftm = torch.bmm(x, trans_feat)
|
|
|
190 |
# GLM-1
|
191 |
sap = torch.bmm(a_s, x_ftm)
|
192 |
sap = sap.transpose(2, 1)
|
@@ -195,12 +168,14 @@ class MeshSegNet(nn.Module):
|
|
195 |
glm_1_sap = F.relu(self.glm1_bn1_2(self.glm1_conv1_2(sap)))
|
196 |
x = torch.cat([x, glm_1_sap], dim=1)
|
197 |
x = F.relu(self.glm1_bn2(self.glm1_conv2(x)))
|
|
|
198 |
# MLP-2
|
199 |
x = F.relu(self.mlp2_bn1(self.mlp2_conv1(x)))
|
200 |
x = F.relu(self.mlp2_bn2(self.mlp2_conv2(x)))
|
201 |
x_mlp2 = F.relu(self.mlp2_bn3(self.mlp2_conv3(x)))
|
202 |
if self.with_dropout:
|
203 |
x_mlp2 = self.dropout(x_mlp2)
|
|
|
204 |
# GLM-2
|
205 |
x_mlp2 = x_mlp2.transpose(2, 1)
|
206 |
sap_1 = torch.bmm(a_s, x_mlp2)
|
@@ -213,12 +188,16 @@ class MeshSegNet(nn.Module):
|
|
213 |
glm_2_sap_2 = F.relu(self.glm2_bn1_3(self.glm2_conv1_3(sap_2)))
|
214 |
x = torch.cat([x, glm_2_sap_1, glm_2_sap_2], dim=1)
|
215 |
x_glm2 = F.relu(self.glm2_bn2(self.glm2_conv2(x)))
|
|
|
216 |
# GMP
|
217 |
x = torch.max(x_glm2, 2, keepdim=True)[0]
|
|
|
218 |
# Upsample
|
219 |
x = torch.nn.Upsample(n_pts)(x)
|
|
|
220 |
# Dense fusion
|
221 |
x = torch.cat([x, x_ftm, x_mlp2, x_glm2], dim=1)
|
|
|
222 |
# MLP-3
|
223 |
x = F.relu(self.mlp3_bn1_1(self.mlp3_conv1(x)))
|
224 |
x = F.relu(self.mlp3_bn1_2(self.mlp3_conv2(x)))
|
@@ -226,6 +205,7 @@ class MeshSegNet(nn.Module):
|
|
226 |
if self.with_dropout:
|
227 |
x = self.dropout(x)
|
228 |
x = F.relu(self.mlp3_bn2_2(self.mlp3_conv4(x)))
|
|
|
229 |
# output
|
230 |
x = self.output_conv(x)
|
231 |
x = x.transpose(2,1).contiguous()
|
@@ -235,20 +215,25 @@ class MeshSegNet(nn.Module):
|
|
235 |
return x
|
236 |
|
237 |
def clone_runoob(li1):
|
|
|
|
|
|
|
238 |
li_copy = li1[:]
|
|
|
239 |
return li_copy
|
240 |
|
241 |
-
#
|
242 |
def class_inlier_outlier(label_list, mean_points,cloud, ind, label_index, points, labels):
|
243 |
label_change = clone_runoob(labels)
|
244 |
outlier_index = clone_runoob(label_index)
|
245 |
ind_reverse = clone_runoob(ind)
|
246 |
-
|
|
|
247 |
ind_reverse.reverse()
|
248 |
for i in ind_reverse:
|
249 |
outlier_index.pop(i)
|
250 |
|
251 |
-
#
|
252 |
inlier_cloud = cloud.select_by_index(ind)
|
253 |
outlier_cloud = cloud.select_by_index(ind, invert=True)
|
254 |
outlier_points = np.array(outlier_cloud.points)
|
@@ -256,29 +241,27 @@ def class_inlier_outlier(label_list, mean_points,cloud, ind, label_index, points
|
|
256 |
for i in range(len(outlier_points)):
|
257 |
distance = []
|
258 |
for j in range(len(mean_points)):
|
259 |
-
dis = np.linalg.norm(outlier_points[i] - mean_points[j], ord=2) #
|
260 |
distance.append(dis)
|
261 |
-
min_index = distance.index(min(distance)) #
|
262 |
-
outlier_label = label_list[min_index] #
|
263 |
index = outlier_index[i]
|
264 |
label_change[index] = outlier_label
|
265 |
|
266 |
return label_change
|
267 |
|
268 |
-
#
|
269 |
def remove_outlier(points, labels):
|
270 |
-
# points = np.array(point_cloud_o3d_orign.points)
|
271 |
-
# global label_list
|
272 |
same_label_points = {}
|
273 |
|
274 |
same_label_index = {}
|
275 |
|
276 |
-
mean_points = []
|
277 |
|
278 |
label_list = []
|
279 |
for i in range(len(labels)):
|
280 |
label_list.append(labels[i])
|
281 |
-
label_list = list(set(label_list))
|
282 |
label_list.sort()
|
283 |
label_list = label_list[1:]
|
284 |
|
@@ -289,7 +272,7 @@ def remove_outlier(points, labels):
|
|
289 |
for j in range(len(labels)):
|
290 |
if labels[j] == i:
|
291 |
points_list.append(points[j].tolist())
|
292 |
-
all_label_index.append(j)
|
293 |
same_label_points[key] = points_list
|
294 |
same_label_index[key] = all_label_index
|
295 |
|
@@ -299,106 +282,102 @@ def remove_outlier(points, labels):
|
|
299 |
|
300 |
for i in label_list:
|
301 |
points_array = same_label_points[i]
|
302 |
-
#
|
303 |
pcd = o3d.geometry.PointCloud()
|
304 |
-
#
|
305 |
pcd.points = o3d.utility.Vector3dVector(points_array)
|
306 |
|
307 |
-
#
|
308 |
-
#
|
309 |
cl, ind = pcd.remove_statistical_outlier(nb_neighbors=200, std_ratio=2.0) # cl是选中的点,ind是选中点index
|
310 |
-
# 可视化
|
311 |
-
# display_inlier_outlier(pcd, ind)
|
312 |
|
313 |
-
#
|
314 |
label_index = same_label_index[i]
|
315 |
labels = class_inlier_outlier(label_list, mean_points, pcd, ind, label_index, points, labels)
|
316 |
# print(f"label_change{labels[4400]}")
|
317 |
|
318 |
return labels
|
319 |
|
320 |
-
|
321 |
-
# 消除离群点,保存最后的输出
|
322 |
def remove_outlier_main(jaw, pcd_points, labels, instances_labels):
|
323 |
-
#
|
324 |
-
# 原始点
|
325 |
points = pcd_points.copy()
|
326 |
label = remove_outlier(points, labels)
|
327 |
|
328 |
-
#
|
329 |
label_dict = {}
|
330 |
label_dict["id_patient"] = ""
|
331 |
label_dict["jaw"] = jaw
|
332 |
label_dict["labels"] = label.tolist()
|
333 |
label_dict["instances"] = instances_labels.tolist()
|
|
|
334 |
b = json.dumps(label_dict)
|
335 |
with open('dental-labels4' + '.json', 'w') as f_obj:
|
336 |
f_obj.write(b)
|
337 |
f_obj.close()
|
338 |
|
339 |
-
|
340 |
same_points_list = {}
|
341 |
|
342 |
-
|
343 |
-
# 体素下采样
|
344 |
def voxel_filter(point_cloud, leaf_size):
|
345 |
same_points_list = {}
|
346 |
filtered_points = []
|
347 |
-
|
|
|
348 |
x_max, y_max, z_max = np.amax(point_cloud, axis=0) # 计算 x,y,z三个维度的最值
|
349 |
x_min, y_min, z_min = np.amin(point_cloud, axis=0)
|
350 |
|
351 |
-
# step2
|
352 |
size_r = leaf_size
|
353 |
|
354 |
-
# step3
|
355 |
Dx = (x_max - x_min) // size_r + 1
|
356 |
Dy = (y_max - y_min) // size_r + 1
|
357 |
Dz = (z_max - z_min) // size_r + 1
|
358 |
|
359 |
# print("Dx x Dy x Dz is {} x {} x {}".format(Dx, Dy, Dz))
|
360 |
|
361 |
-
# step4
|
362 |
-
h = list() # h
|
363 |
for i in range(len(point_cloud)):
|
364 |
hx = np.floor((point_cloud[i][0] - x_min) // size_r)
|
365 |
hy = np.floor((point_cloud[i][1] - y_min) // size_r)
|
366 |
hz = np.floor((point_cloud[i][2] - z_min) // size_r)
|
367 |
h.append(hx + hy * Dx + hz * Dx * Dy)
|
368 |
-
# print(h[60581])
|
369 |
|
370 |
-
# step5
|
371 |
h = np.array(h)
|
372 |
-
h_indice = np.argsort(h) #
|
373 |
-
h_sorted = h[h_indice] #
|
374 |
-
count = 0 #
|
375 |
step = 20
|
376 |
-
|
377 |
-
|
378 |
-
|
379 |
-
# print("aaa")
|
380 |
if h_sorted[i] == h_sorted[i - 1] and (i != len(h_sorted) - 1):
|
381 |
continue
|
|
|
382 |
elif h_sorted[i] == h_sorted[i - 1] and (i == len(h_sorted) - 1):
|
383 |
point_idx = h_indice[count:]
|
384 |
key = h_sorted[i - 1]
|
385 |
same_points_list[key] = point_idx
|
386 |
-
_G = np.mean(point_cloud[point_idx], axis=0) #
|
387 |
-
_d = np.linalg.norm(point_cloud[point_idx] - _G, axis=1, ord=2) #
|
388 |
_d.sort()
|
389 |
-
inx = [j for j in range(0, len(_d), step)] #
|
390 |
for j in inx:
|
391 |
index = point_idx[j]
|
392 |
filtered_points.append(point_cloud[index])
|
393 |
count = i
|
|
|
394 |
elif h_sorted[i] != h_sorted[i - 1] and (i == len(h_sorted) - 1):
|
395 |
point_idx1 = h_indice[count:i]
|
396 |
key1 = h_sorted[i - 1]
|
397 |
same_points_list[key1] = point_idx1
|
398 |
-
_G = np.mean(point_cloud[point_idx1], axis=0) #
|
399 |
-
_d = np.linalg.norm(point_cloud[point_idx1] - _G, axis=1, ord=2) #
|
400 |
_d.sort()
|
401 |
-
inx = [j for j in range(0, len(_d), step)] #
|
402 |
for j in inx:
|
403 |
index = point_idx1[j]
|
404 |
filtered_points.append(point_cloud[index])
|
@@ -406,10 +385,10 @@ def voxel_filter(point_cloud, leaf_size):
|
|
406 |
point_idx2 = h_indice[i:]
|
407 |
key2 = h_sorted[i]
|
408 |
same_points_list[key2] = point_idx2
|
409 |
-
_G = np.mean(point_cloud[point_idx2], axis=0) #
|
410 |
-
_d = np.linalg.norm(point_cloud[point_idx2] - _G, axis=1, ord=2) #
|
411 |
_d.sort()
|
412 |
-
inx = [j for j in range(0, len(_d), step)] #
|
413 |
for j in inx:
|
414 |
index = point_idx2[j]
|
415 |
filtered_points.append(point_cloud[index])
|
@@ -419,38 +398,42 @@ def voxel_filter(point_cloud, leaf_size):
|
|
419 |
point_idx = h_indice[count: i]
|
420 |
key = h_sorted[i - 1]
|
421 |
same_points_list[key] = point_idx
|
422 |
-
_G = np.mean(point_cloud[point_idx], axis=0) #
|
423 |
-
_d = np.linalg.norm(point_cloud[point_idx] - _G, axis=1, ord=2) #
|
424 |
_d.sort()
|
425 |
-
inx = [j for j in range(0, len(_d), step)] #
|
426 |
for j in inx:
|
427 |
index = point_idx[j]
|
428 |
filtered_points.append(point_cloud[index])
|
429 |
count = i
|
430 |
|
431 |
-
#
|
432 |
# print(f'filtered_points[0]为{filtered_points[0]}')
|
433 |
filtered_points = np.array(filtered_points, dtype=np.float64)
|
|
|
434 |
return filtered_points,same_points_list
|
435 |
|
436 |
|
437 |
-
#
|
438 |
def voxel_upsample(same_points_list, point_cloud, filtered_points, filter_labels, leaf_size):
|
439 |
upsample_label = []
|
440 |
upsample_point = []
|
441 |
upsample_index = []
|
442 |
-
|
443 |
-
|
|
|
444 |
x_min, y_min, z_min = np.amin(point_cloud, axis=0)
|
445 |
-
|
|
|
446 |
size_r = leaf_size
|
447 |
-
|
|
|
448 |
Dx = (x_max - x_min) // size_r + 1
|
449 |
Dy = (y_max - y_min) // size_r + 1
|
450 |
Dz = (z_max - z_min) // size_r + 1
|
451 |
print("Dx x Dy x Dz is {} x {} x {}".format(Dx, Dy, Dz))
|
452 |
|
453 |
-
# step4
|
454 |
h = list()
|
455 |
for i in range(len(filtered_points)):
|
456 |
hx = np.floor((filtered_points[i][0] - x_min) // size_r)
|
@@ -458,30 +441,33 @@ def voxel_upsample(same_points_list, point_cloud, filtered_points, filter_labels
|
|
458 |
hz = np.floor((filtered_points[i][2] - z_min) // size_r)
|
459 |
h.append(hx + hy * Dx + hz * Dx * Dy)
|
460 |
|
461 |
-
# step5
|
462 |
h = np.array(h)
|
463 |
count = 0
|
464 |
for i in range(1, len(h)):
|
465 |
if h[i] == h[i - 1] and i != (len(h) - 1):
|
466 |
continue
|
|
|
467 |
elif h[i] == h[i - 1] and i == (len(h) - 1):
|
468 |
label = filter_labels[count:]
|
469 |
key = h[i - 1]
|
470 |
count = i
|
471 |
-
|
|
|
472 |
classcount = {}
|
473 |
for i in range(len(label)):
|
474 |
vote = label[i]
|
475 |
classcount[vote] = classcount.get(vote, 0) + 1
|
476 |
-
|
|
|
477 |
sortedclass = sorted(classcount.items(), key=lambda x: (x[1]), reverse=True)
|
478 |
-
|
479 |
-
point_index = same_points_list[key] # h对应的point index列表
|
480 |
for j in range(len(point_index)):
|
481 |
upsample_label.append(sortedclass[0][0])
|
482 |
index = point_index[j]
|
483 |
upsample_point.append(point_cloud[index])
|
484 |
upsample_index.append(index)
|
|
|
485 |
elif h[i] != h[i - 1] and (i == len(h) - 1):
|
486 |
label1 = filter_labels[count:i]
|
487 |
key1 = h[i - 1]
|
@@ -493,8 +479,8 @@ def voxel_upsample(same_points_list, point_cloud, filtered_points, filter_labels
|
|
493 |
for i in range(len(label1)):
|
494 |
vote = label1[i]
|
495 |
classcount[vote] = classcount.get(vote, 0) + 1
|
|
|
496 |
sortedclass = sorted(classcount.items(), key=lambda x: (x[1]), reverse=True)
|
497 |
-
# key1 = h[i-1]
|
498 |
point_index = same_points_list[key1]
|
499 |
for j in range(len(point_index)):
|
500 |
upsample_label.append(sortedclass[0][0])
|
@@ -502,13 +488,12 @@ def voxel_upsample(same_points_list, point_cloud, filtered_points, filter_labels
|
|
502 |
upsample_point.append(point_cloud[index])
|
503 |
upsample_index.append(index)
|
504 |
|
505 |
-
# label2 = filter_labels[i:]
|
506 |
classcount = {}
|
507 |
for i in range(len(label2)):
|
508 |
vote = label2[i]
|
509 |
classcount[vote] = classcount.get(vote, 0) + 1
|
|
|
510 |
sortedclass = sorted(classcount.items(), key=lambda x: (x[1]), reverse=True)
|
511 |
-
# key2 = h[i]
|
512 |
point_index = same_points_list[key2]
|
513 |
for j in range(len(point_index)):
|
514 |
upsample_label.append(sortedclass[0][0])
|
@@ -523,58 +508,51 @@ def voxel_upsample(same_points_list, point_cloud, filtered_points, filter_labels
|
|
523 |
for i in range(len(label)):
|
524 |
vote = label[i]
|
525 |
classcount[vote] = classcount.get(vote, 0) + 1
|
|
|
526 |
sortedclass = sorted(classcount.items(), key=lambda x: (x[1]), reverse=True)
|
527 |
-
# key = h[i-1]
|
528 |
point_index = same_points_list[key] # h对应的point index列表
|
529 |
for j in range(len(point_index)):
|
530 |
upsample_label.append(sortedclass[0][0])
|
531 |
index = point_index[j]
|
532 |
upsample_point.append(point_cloud[index])
|
533 |
upsample_index.append(index)
|
534 |
-
# count = i
|
535 |
|
536 |
-
#
|
537 |
-
# print(f'upsample_index[0]的值为{upsample_index[0]}')
|
538 |
-
# print(f'upsample_index的总长度为{len(upsample_index)}')
|
539 |
-
|
540 |
-
# 恢复index原始顺序
|
541 |
upsample_index = np.array(upsample_index)
|
542 |
-
upsample_index_indice = np.argsort(upsample_index)
|
543 |
upsample_index_sorted = upsample_index[upsample_index_indice]
|
544 |
|
545 |
upsample_point = np.array(upsample_point)
|
546 |
upsample_label = np.array(upsample_label)
|
547 |
-
|
|
|
548 |
upsample_point_sorted = upsample_point[upsample_index_indice]
|
549 |
upsample_label_sorted = upsample_label[upsample_index_indice]
|
550 |
|
551 |
return upsample_point_sorted, upsample_label_sorted
|
552 |
|
553 |
-
|
554 |
-
# 利用knn算法上采样
|
555 |
def KNN_sklearn_Load_data(voxel_points, center_points, labels):
|
556 |
-
#
|
557 |
-
# x_train, x_test, y_train, y_test = train_test_split(center_points, labels, test_size=0.1)
|
558 |
-
# 构建模型
|
559 |
model = neighbors.KNeighborsClassifier(n_neighbors=3)
|
560 |
model.fit(center_points, labels)
|
561 |
prediction = model.predict(voxel_points.reshape(1, -1))
|
562 |
-
# meshtopoints_labels = classification_report(voxel_points, prediction)
|
563 |
-
return prediction[0]
|
564 |
|
|
|
565 |
|
566 |
-
#
|
567 |
def Load_data(voxel_points, center_points, labels):
|
568 |
meshtopoints_labels = []
|
569 |
-
# meshtopoints_labels.append(SVC_sklearn_Load_data(voxel_points[i], center_points, labels))
|
570 |
for i in range(0, voxel_points.shape[0]):
|
571 |
meshtopoints_labels.append(KNN_sklearn_Load_data(voxel_points[i], center_points, labels))
|
|
|
572 |
return np.array(meshtopoints_labels)
|
573 |
|
574 |
-
#
|
575 |
def mesh_to_points_main(jaw, pcd_points, center_points, labels):
|
576 |
points = pcd_points.copy()
|
577 |
-
|
|
|
578 |
voxel_points, same_points_list = voxel_filter(points, 0.6)
|
579 |
|
580 |
after_labels = Load_data(voxel_points, center_points, labels)
|
@@ -584,8 +562,8 @@ def mesh_to_points_main(jaw, pcd_points, center_points, labels):
|
|
584 |
new_pcd = o3d.geometry.PointCloud()
|
585 |
new_pcd.points = o3d.utility.Vector3dVector(upsample_point)
|
586 |
instances_labels = upsample_label.copy()
|
587 |
-
|
588 |
-
#
|
589 |
for i in stqdm(range(0, upsample_label.shape[0])):
|
590 |
if jaw == 'upper':
|
591 |
if (upsample_label[i] >= 1) and (upsample_label[i] <= 8):
|
@@ -597,13 +575,14 @@ def mesh_to_points_main(jaw, pcd_points, center_points, labels):
|
|
597 |
upsample_label[i] = upsample_label[i] + 30
|
598 |
elif (upsample_label[i] >= 9) and (upsample_label[i] <= 16):
|
599 |
upsample_label[i] = upsample_label[i] + 32
|
|
|
600 |
remove_outlier_main(jaw, pcd_points, upsample_label, instances_labels)
|
601 |
|
602 |
|
603 |
-
#
|
604 |
def mesh_grid(pcd_points):
|
605 |
new_pcd,_ = voxel_filter(pcd_points, 0.6)
|
606 |
-
# pcd
|
607 |
|
608 |
# estimate radius for rolling ball
|
609 |
pcd_new = o3d.geometry.PointCloud()
|
@@ -615,12 +594,10 @@ def mesh_grid(pcd_points):
|
|
615 |
mesh = o3d.geometry.TriangleMesh.create_from_point_cloud_ball_pivoting(
|
616 |
pcd_new,
|
617 |
o3d.utility.DoubleVector([radius, radius * 2]))
|
618 |
-
# o3d.io.write_triangle_mesh("./tooth date/test.ply", mesh)
|
619 |
|
620 |
return mesh
|
621 |
|
622 |
-
|
623 |
-
# 读取obj文件内容
|
624 |
def read_obj(obj_path):
|
625 |
jaw = None
|
626 |
with open(obj_path) as file:
|
@@ -642,14 +619,12 @@ def read_obj(obj_path):
|
|
642 |
|
643 |
points = np.array(points)
|
644 |
faces = np.array(faces)
|
645 |
-
|
646 |
if jaw is None:
|
647 |
raise ValueError("Jaw type not found in OBJ file")
|
648 |
|
649 |
return points, faces, jaw
|
650 |
|
651 |
-
|
652 |
-
# obj文件转为pcd文件
|
653 |
def obj2pcd(obj_path):
|
654 |
if os.path.exists(obj_path):
|
655 |
print('yes')
|
@@ -661,13 +636,14 @@ def obj2pcd(obj_path):
|
|
661 |
pcd_list.append(new_line.split())
|
662 |
|
663 |
pcd_points = np.array(pcd_list).astype(np.float64)
|
664 |
-
return pcd_points, jaw
|
665 |
|
|
|
666 |
|
|
|
667 |
def segmentation_main(obj_path):
|
668 |
upsampling_method = 'KNN'
|
669 |
|
670 |
-
model_path = '
|
671 |
num_classes = 17
|
672 |
num_channels = 15
|
673 |
|
@@ -737,6 +713,7 @@ def segmentation_main(obj_path):
|
|
737 |
nmeans = normals.mean(axis=0)
|
738 |
nstds = normals.std(axis=0)
|
739 |
|
|
|
740 |
for i in range(3):
|
741 |
cells[:, i] = (cells[:, i] - means[i]) / stds[i] # point 1
|
742 |
cells[:, i + 3] = (cells[:, i + 3] - means[i]) / stds[i] # point 2
|
@@ -744,6 +721,7 @@ def segmentation_main(obj_path):
|
|
744 |
barycenters[:, i] = (barycenters[:, i] - mins[i]) / (maxs[i] - mins[i])
|
745 |
normals[:, i] = (normals[:, i] - nmeans[i]) / nstds[i]
|
746 |
|
|
|
747 |
X = np.column_stack((cells, barycenters, normals))
|
748 |
|
749 |
# computing A_S and A_L
|
@@ -794,6 +772,7 @@ def segmentation_main(obj_path):
|
|
794 |
if i_node < i_nei:
|
795 |
cos_theta = np.dot(normals[i_node, 0:3], normals[i_nei, 0:3]) / np.linalg.norm(
|
796 |
normals[i_node, 0:3]) / np.linalg.norm(normals[i_nei, 0:3])
|
|
|
797 |
if cos_theta >= 1.0:
|
798 |
cos_theta = 0.9999
|
799 |
theta = np.arccos(cos_theta)
|
@@ -806,6 +785,7 @@ def segmentation_main(obj_path):
|
|
806 |
edges = np.concatenate(
|
807 |
(edges, np.array([i_node, i_nei, -beta * np.log10(theta / np.pi) * phi]).reshape(1, 3)),
|
808 |
axis=0)
|
|
|
809 |
edges = np.delete(edges, 0, 0)
|
810 |
edges[:, 2] *= lambda_c * round_factor
|
811 |
edges = edges.astype(np.int32)
|
@@ -913,9 +893,9 @@ class Segment(TeethApp):
|
|
913 |
# Create a pyvista plotter
|
914 |
plotter = pv.Plotter()
|
915 |
|
916 |
-
cmap = plt.cm.get_cmap('jet', 27)
|
917 |
|
918 |
-
colors = cmap(np.linspace(0, 1, 27))
|
919 |
|
920 |
# Convert colors to a format acceptable by PyVista
|
921 |
colormap = mcolors.ListedColormap(colors)
|
@@ -930,8 +910,6 @@ class Segment(TeethApp):
|
|
930 |
with st.expander("Ground Truth - scroll for zoom", expanded=False):
|
931 |
stpyvista(plotter)
|
932 |
|
933 |
-
|
934 |
-
|
935 |
elif inputs == "Upload Scan":
|
936 |
file = st.file_uploader("Please upload an OBJ Object file", type=["OBJ"])
|
937 |
st.markdown("Expected time per prediction: 7-10 min.")
|
@@ -939,7 +917,7 @@ class Segment(TeethApp):
|
|
939 |
# save the uploaded file to disk
|
940 |
with open("file.obj", "wb") as buffer:
|
941 |
shutil.copyfileobj(file, buffer)
|
942 |
-
|
943 |
obj_path = "file.obj"
|
944 |
|
945 |
mesh = pv.read(obj_path)
|
@@ -957,9 +935,5 @@ class Segment(TeethApp):
|
|
957 |
if segment:
|
958 |
segmentation_main(obj_path)
|
959 |
|
960 |
-
|
961 |
-
|
962 |
-
|
963 |
-
|
964 |
if __name__ == "__main__":
|
965 |
app = Segment()
|
|
|
1 |
+
import os
|
2 |
import shutil
|
3 |
+
import json
|
4 |
|
|
|
5 |
import numpy as np
|
|
|
6 |
from scipy.spatial import distance_matrix
|
7 |
+
from sklearn import neighbors
|
8 |
from pygco import cut_from_graph
|
9 |
import open3d as o3d
|
10 |
import matplotlib.pyplot as plt
|
11 |
import matplotlib.colors as mcolors
|
|
|
|
|
|
|
12 |
import torch
|
13 |
import torch.nn as nn
|
14 |
from torch.autograd import Variable
|
15 |
import torch.nn.functional as F
|
16 |
import streamlit as st
|
17 |
+
from streamlit import session_state as session
|
18 |
+
from stpyvista import stpyvista
|
19 |
+
from stqdm import stqdm
|
20 |
from PIL import Image
|
21 |
|
22 |
+
# Configure Streamlit page
|
23 |
class TeethApp:
|
24 |
+
"""
|
25 |
+
Base class for Streamlit app
|
26 |
+
"""
|
27 |
def __init__(self):
|
28 |
# Font
|
29 |
with open("utils/style.css") as css:
|
|
|
50 |
unsafe_allow_html=True,
|
51 |
)
|
52 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
53 |
class STNkd(nn.Module):
|
54 |
def __init__(self, k=64):
|
55 |
super(STNkd, self).__init__()
|
|
|
97 |
self.with_dropout = with_dropout
|
98 |
self.dropout_p = dropout_p
|
99 |
|
100 |
+
# MLP-1 -shape: [64, 64]
|
101 |
self.mlp1_conv1 = torch.nn.Conv1d(self.num_channels, 64, 1)
|
102 |
self.mlp1_conv2 = torch.nn.Conv1d(64, 64, 1)
|
103 |
self.mlp1_bn1 = nn.BatchNorm1d(64)
|
104 |
self.mlp1_bn2 = nn.BatchNorm1d(64)
|
105 |
+
|
106 |
# FTM (feature-transformer module)
|
107 |
self.fstn = STNkd(k=64)
|
108 |
+
|
109 |
# GLM-1 (graph-contrained learning modulus)
|
110 |
self.glm1_conv1_1 = torch.nn.Conv1d(64, 32, 1)
|
111 |
self.glm1_conv1_2 = torch.nn.Conv1d(64, 32, 1)
|
|
|
113 |
self.glm1_bn1_2 = nn.BatchNorm1d(32)
|
114 |
self.glm1_conv2 = torch.nn.Conv1d(32+32, 64, 1)
|
115 |
self.glm1_bn2 = nn.BatchNorm1d(64)
|
116 |
+
|
117 |
# MLP-2
|
118 |
self.mlp2_conv1 = torch.nn.Conv1d(64, 64, 1)
|
119 |
self.mlp2_bn1 = nn.BatchNorm1d(64)
|
|
|
121 |
self.mlp2_bn2 = nn.BatchNorm1d(128)
|
122 |
self.mlp2_conv3 = torch.nn.Conv1d(128, 512, 1)
|
123 |
self.mlp2_bn3 = nn.BatchNorm1d(512)
|
124 |
+
|
125 |
# GLM-2 (graph-contrained learning modulus)
|
126 |
self.glm2_conv1_1 = torch.nn.Conv1d(512, 128, 1)
|
127 |
self.glm2_conv1_2 = torch.nn.Conv1d(512, 128, 1)
|
|
|
131 |
self.glm2_bn1_3 = nn.BatchNorm1d(128)
|
132 |
self.glm2_conv2 = torch.nn.Conv1d(128*3, 512, 1)
|
133 |
self.glm2_bn2 = nn.BatchNorm1d(512)
|
134 |
+
|
135 |
# MLP-3
|
136 |
self.mlp3_conv1 = torch.nn.Conv1d(64+512+512+512, 256, 1)
|
137 |
self.mlp3_conv2 = torch.nn.Conv1d(256, 256, 1)
|
|
|
141 |
self.mlp3_conv4 = torch.nn.Conv1d(128, 128, 1)
|
142 |
self.mlp3_bn2_1 = nn.BatchNorm1d(128)
|
143 |
self.mlp3_bn2_2 = nn.BatchNorm1d(128)
|
144 |
+
|
145 |
+
# Output
|
146 |
self.output_conv = torch.nn.Conv1d(128, self.num_classes, 1)
|
147 |
if self.with_dropout:
|
148 |
self.dropout = nn.Dropout(p=self.dropout_p)
|
|
|
150 |
def forward(self, x, a_s, a_l):
|
151 |
batchsize = x.size()[0]
|
152 |
n_pts = x.size()[2]
|
153 |
+
|
154 |
# MLP-1
|
155 |
x = F.relu(self.mlp1_bn1(self.mlp1_conv1(x)))
|
156 |
x = F.relu(self.mlp1_bn2(self.mlp1_conv2(x)))
|
157 |
+
|
158 |
# FTM
|
159 |
trans_feat = self.fstn(x)
|
160 |
x = x.transpose(2, 1)
|
161 |
x_ftm = torch.bmm(x, trans_feat)
|
162 |
+
|
163 |
# GLM-1
|
164 |
sap = torch.bmm(a_s, x_ftm)
|
165 |
sap = sap.transpose(2, 1)
|
|
|
168 |
glm_1_sap = F.relu(self.glm1_bn1_2(self.glm1_conv1_2(sap)))
|
169 |
x = torch.cat([x, glm_1_sap], dim=1)
|
170 |
x = F.relu(self.glm1_bn2(self.glm1_conv2(x)))
|
171 |
+
|
172 |
# MLP-2
|
173 |
x = F.relu(self.mlp2_bn1(self.mlp2_conv1(x)))
|
174 |
x = F.relu(self.mlp2_bn2(self.mlp2_conv2(x)))
|
175 |
x_mlp2 = F.relu(self.mlp2_bn3(self.mlp2_conv3(x)))
|
176 |
if self.with_dropout:
|
177 |
x_mlp2 = self.dropout(x_mlp2)
|
178 |
+
|
179 |
# GLM-2
|
180 |
x_mlp2 = x_mlp2.transpose(2, 1)
|
181 |
sap_1 = torch.bmm(a_s, x_mlp2)
|
|
|
188 |
glm_2_sap_2 = F.relu(self.glm2_bn1_3(self.glm2_conv1_3(sap_2)))
|
189 |
x = torch.cat([x, glm_2_sap_1, glm_2_sap_2], dim=1)
|
190 |
x_glm2 = F.relu(self.glm2_bn2(self.glm2_conv2(x)))
|
191 |
+
|
192 |
# GMP
|
193 |
x = torch.max(x_glm2, 2, keepdim=True)[0]
|
194 |
+
|
195 |
# Upsample
|
196 |
x = torch.nn.Upsample(n_pts)(x)
|
197 |
+
|
198 |
# Dense fusion
|
199 |
x = torch.cat([x, x_ftm, x_mlp2, x_glm2], dim=1)
|
200 |
+
|
201 |
# MLP-3
|
202 |
x = F.relu(self.mlp3_bn1_1(self.mlp3_conv1(x)))
|
203 |
x = F.relu(self.mlp3_bn1_2(self.mlp3_conv2(x)))
|
|
|
205 |
if self.with_dropout:
|
206 |
x = self.dropout(x)
|
207 |
x = F.relu(self.mlp3_bn2_2(self.mlp3_conv4(x)))
|
208 |
+
|
209 |
# output
|
210 |
x = self.output_conv(x)
|
211 |
x = x.transpose(2,1).contiguous()
|
|
|
215 |
return x
|
216 |
|
217 |
def clone_runoob(li1):
|
218 |
+
"""
|
219 |
+
copy list
|
220 |
+
"""
|
221 |
li_copy = li1[:]
|
222 |
+
|
223 |
return li_copy
|
224 |
|
225 |
+
# Reclassify outliers
|
226 |
def class_inlier_outlier(label_list, mean_points,cloud, ind, label_index, points, labels):
|
227 |
label_change = clone_runoob(labels)
|
228 |
outlier_index = clone_runoob(label_index)
|
229 |
ind_reverse = clone_runoob(ind)
|
230 |
+
|
231 |
+
# Get the label subscript of the outlier point
|
232 |
ind_reverse.reverse()
|
233 |
for i in ind_reverse:
|
234 |
outlier_index.pop(i)
|
235 |
|
236 |
+
# Get outliers
|
237 |
inlier_cloud = cloud.select_by_index(ind)
|
238 |
outlier_cloud = cloud.select_by_index(ind, invert=True)
|
239 |
outlier_points = np.array(outlier_cloud.points)
|
|
|
241 |
for i in range(len(outlier_points)):
|
242 |
distance = []
|
243 |
for j in range(len(mean_points)):
|
244 |
+
dis = np.linalg.norm(outlier_points[i] - mean_points[j], ord=2) # Compute the distance between tooth and GT centroid
|
245 |
distance.append(dis)
|
246 |
+
min_index = distance.index(min(distance)) # Get the index of the label closest to the centroid of the outlier point
|
247 |
+
outlier_label = label_list[min_index] # Get the label of the outlier point
|
248 |
index = outlier_index[i]
|
249 |
label_change[index] = outlier_label
|
250 |
|
251 |
return label_change
|
252 |
|
253 |
+
# Use knn algorithm to eliminate outliers
|
254 |
def remove_outlier(points, labels):
|
|
|
|
|
255 |
same_label_points = {}
|
256 |
|
257 |
same_label_index = {}
|
258 |
|
259 |
+
mean_points = [] # All label types correspond to the centroid coordinates of the point cloud.
|
260 |
|
261 |
label_list = []
|
262 |
for i in range(len(labels)):
|
263 |
label_list.append(labels[i])
|
264 |
+
label_list = list(set(label_list)) # To retrieve the order from small to large, take GT_label=[0, 11, 12, 13, 14, 15, 16, 17, 21, 22, 23, 24, 25, 26, 27]
|
265 |
label_list.sort()
|
266 |
label_list = label_list[1:]
|
267 |
|
|
|
272 |
for j in range(len(labels)):
|
273 |
if labels[j] == i:
|
274 |
points_list.append(points[j].tolist())
|
275 |
+
all_label_index.append(j) # Get the subscript of the label corresponding to the point with label i
|
276 |
same_label_points[key] = points_list
|
277 |
same_label_index[key] = all_label_index
|
278 |
|
|
|
282 |
|
283 |
for i in label_list:
|
284 |
points_array = same_label_points[i]
|
285 |
+
# Build one o3d object
|
286 |
pcd = o3d.geometry.PointCloud()
|
287 |
+
# UseVector3dVector conversion method
|
288 |
pcd.points = o3d.utility.Vector3dVector(points_array)
|
289 |
|
290 |
+
# Perform statistical outlier removal on the point cloud corresponding to label i, find outliers and display them
|
291 |
+
# Statistical outlier removal
|
292 |
cl, ind = pcd.remove_statistical_outlier(nb_neighbors=200, std_ratio=2.0) # cl是选中的点,ind是选中点index
|
|
|
|
|
293 |
|
294 |
+
# Reclassify the separated outliers
|
295 |
label_index = same_label_index[i]
|
296 |
labels = class_inlier_outlier(label_list, mean_points, pcd, ind, label_index, points, labels)
|
297 |
# print(f"label_change{labels[4400]}")
|
298 |
|
299 |
return labels
|
300 |
|
301 |
+
# Eliminate outliers and save the final output
|
|
|
302 |
def remove_outlier_main(jaw, pcd_points, labels, instances_labels):
|
303 |
+
# original point
|
|
|
304 |
points = pcd_points.copy()
|
305 |
label = remove_outlier(points, labels)
|
306 |
|
307 |
+
# Save json file
|
308 |
label_dict = {}
|
309 |
label_dict["id_patient"] = ""
|
310 |
label_dict["jaw"] = jaw
|
311 |
label_dict["labels"] = label.tolist()
|
312 |
label_dict["instances"] = instances_labels.tolist()
|
313 |
+
|
314 |
b = json.dumps(label_dict)
|
315 |
with open('dental-labels4' + '.json', 'w') as f_obj:
|
316 |
f_obj.write(b)
|
317 |
f_obj.close()
|
318 |
|
|
|
319 |
same_points_list = {}
|
320 |
|
321 |
+
# voxel downsampling
|
|
|
322 |
def voxel_filter(point_cloud, leaf_size):
|
323 |
same_points_list = {}
|
324 |
filtered_points = []
|
325 |
+
|
326 |
+
# step1 Calculate boundary points
|
327 |
x_max, y_max, z_max = np.amax(point_cloud, axis=0) # 计算 x,y,z三个维度的最值
|
328 |
x_min, y_min, z_min = np.amin(point_cloud, axis=0)
|
329 |
|
330 |
+
# step2 Determine the size of the voxel
|
331 |
size_r = leaf_size
|
332 |
|
333 |
+
# step3 Calculate the dimensions of each volex voxel grid
|
334 |
Dx = (x_max - x_min) // size_r + 1
|
335 |
Dy = (y_max - y_min) // size_r + 1
|
336 |
Dz = (z_max - z_min) // size_r + 1
|
337 |
|
338 |
# print("Dx x Dy x Dz is {} x {} x {}".format(Dx, Dy, Dz))
|
339 |
|
340 |
+
# step4 Calculate the value of each point in each dimension in the volex grid
|
341 |
+
h = list() # h is a list of saved indexes
|
342 |
for i in range(len(point_cloud)):
|
343 |
hx = np.floor((point_cloud[i][0] - x_min) // size_r)
|
344 |
hy = np.floor((point_cloud[i][1] - y_min) // size_r)
|
345 |
hz = np.floor((point_cloud[i][2] - z_min) // size_r)
|
346 |
h.append(hx + hy * Dx + hz * Dx * Dy)
|
|
|
347 |
|
348 |
+
# step5 Sort h values
|
349 |
h = np.array(h)
|
350 |
+
h_indice = np.argsort(h) # Extract the index and return the index of the elements in h sorted from small to large.
|
351 |
+
h_sorted = h[h_indice] # Ascending order
|
352 |
+
count = 0 # used for accumulation of dimensions
|
353 |
step = 20
|
354 |
+
|
355 |
+
# Put points with the same h value into the same grid and filter them
|
356 |
+
for i in range(1, len(h_sorted)): # 0-19999 data points
|
|
|
357 |
if h_sorted[i] == h_sorted[i - 1] and (i != len(h_sorted) - 1):
|
358 |
continue
|
359 |
+
|
360 |
elif h_sorted[i] == h_sorted[i - 1] and (i == len(h_sorted) - 1):
|
361 |
point_idx = h_indice[count:]
|
362 |
key = h_sorted[i - 1]
|
363 |
same_points_list[key] = point_idx
|
364 |
+
_G = np.mean(point_cloud[point_idx], axis=0) # center of gravity of all points
|
365 |
+
_d = np.linalg.norm(point_cloud[point_idx] - _G, axis=1, ord=2) # Calculate distance to center of gravity
|
366 |
_d.sort()
|
367 |
+
inx = [j for j in range(0, len(_d), step)] # Get the index of the specified interval element
|
368 |
for j in inx:
|
369 |
index = point_idx[j]
|
370 |
filtered_points.append(point_cloud[index])
|
371 |
count = i
|
372 |
+
|
373 |
elif h_sorted[i] != h_sorted[i - 1] and (i == len(h_sorted) - 1):
|
374 |
point_idx1 = h_indice[count:i]
|
375 |
key1 = h_sorted[i - 1]
|
376 |
same_points_list[key1] = point_idx1
|
377 |
+
_G = np.mean(point_cloud[point_idx1], axis=0) # center of gravity of all points
|
378 |
+
_d = np.linalg.norm(point_cloud[point_idx1] - _G, axis=1, ord=2) # Calculate distance to center of gravity
|
379 |
_d.sort()
|
380 |
+
inx = [j for j in range(0, len(_d), step)] # Get the index of the specified interval element
|
381 |
for j in inx:
|
382 |
index = point_idx1[j]
|
383 |
filtered_points.append(point_cloud[index])
|
|
|
385 |
point_idx2 = h_indice[i:]
|
386 |
key2 = h_sorted[i]
|
387 |
same_points_list[key2] = point_idx2
|
388 |
+
_G = np.mean(point_cloud[point_idx2], axis=0) # center of gravity of all points
|
389 |
+
_d = np.linalg.norm(point_cloud[point_idx2] - _G, axis=1, ord=2) # Calculate distance to center of gravity
|
390 |
_d.sort()
|
391 |
+
inx = [j for j in range(0, len(_d), step)] # Get the index of the specified interval element
|
392 |
for j in inx:
|
393 |
index = point_idx2[j]
|
394 |
filtered_points.append(point_cloud[index])
|
|
|
398 |
point_idx = h_indice[count: i]
|
399 |
key = h_sorted[i - 1]
|
400 |
same_points_list[key] = point_idx
|
401 |
+
_G = np.mean(point_cloud[point_idx], axis=0) # center of gravity of all points
|
402 |
+
_d = np.linalg.norm(point_cloud[point_idx] - _G, axis=1, ord=2) # Calculate distance to center of gravity
|
403 |
_d.sort()
|
404 |
+
inx = [j for j in range(0, len(_d), step)] # Get the index of the specified interval element
|
405 |
for j in inx:
|
406 |
index = point_idx[j]
|
407 |
filtered_points.append(point_cloud[index])
|
408 |
count = i
|
409 |
|
410 |
+
# Change the point cloud format to array and return it externally
|
411 |
# print(f'filtered_points[0]为{filtered_points[0]}')
|
412 |
filtered_points = np.array(filtered_points, dtype=np.float64)
|
413 |
+
|
414 |
return filtered_points,same_points_list
|
415 |
|
416 |
|
417 |
+
# voxel upsampling
|
418 |
def voxel_upsample(same_points_list, point_cloud, filtered_points, filter_labels, leaf_size):
|
419 |
upsample_label = []
|
420 |
upsample_point = []
|
421 |
upsample_index = []
|
422 |
+
|
423 |
+
# step1 Calculate boundary points
|
424 |
+
x_max, y_max, z_max = np.amax(point_cloud, axis=0) # Calculate the maximum value of the three dimensions x, y, z
|
425 |
x_min, y_min, z_min = np.amin(point_cloud, axis=0)
|
426 |
+
|
427 |
+
# step2 Determine the size of the voxel
|
428 |
size_r = leaf_size
|
429 |
+
|
430 |
+
# step3 Calculate the dimensions of each volex voxel grid
|
431 |
Dx = (x_max - x_min) // size_r + 1
|
432 |
Dy = (y_max - y_min) // size_r + 1
|
433 |
Dz = (z_max - z_min) // size_r + 1
|
434 |
print("Dx x Dy x Dz is {} x {} x {}".format(Dx, Dy, Dz))
|
435 |
|
436 |
+
# step4 Calculate the value of each point (sampled point) in each dimension within the volex grid
|
437 |
h = list()
|
438 |
for i in range(len(filtered_points)):
|
439 |
hx = np.floor((filtered_points[i][0] - x_min) // size_r)
|
|
|
441 |
hz = np.floor((filtered_points[i][2] - z_min) // size_r)
|
442 |
h.append(hx + hy * Dx + hz * Dx * Dy)
|
443 |
|
444 |
+
# step5 Query the dictionary same_points_list based on the h value
|
445 |
h = np.array(h)
|
446 |
count = 0
|
447 |
for i in range(1, len(h)):
|
448 |
if h[i] == h[i - 1] and i != (len(h) - 1):
|
449 |
continue
|
450 |
+
|
451 |
elif h[i] == h[i - 1] and i == (len(h) - 1):
|
452 |
label = filter_labels[count:]
|
453 |
key = h[i - 1]
|
454 |
count = i
|
455 |
+
|
456 |
+
# Cumulative number of labels, classcount: {‘A’: 2, ‘B’: 1}
|
457 |
classcount = {}
|
458 |
for i in range(len(label)):
|
459 |
vote = label[i]
|
460 |
classcount[vote] = classcount.get(vote, 0) + 1
|
461 |
+
|
462 |
+
# Sort map values
|
463 |
sortedclass = sorted(classcount.items(), key=lambda x: (x[1]), reverse=True)
|
464 |
+
point_index = same_points_list[key] # Point index list corresponding to h
|
|
|
465 |
for j in range(len(point_index)):
|
466 |
upsample_label.append(sortedclass[0][0])
|
467 |
index = point_index[j]
|
468 |
upsample_point.append(point_cloud[index])
|
469 |
upsample_index.append(index)
|
470 |
+
|
471 |
elif h[i] != h[i - 1] and (i == len(h) - 1):
|
472 |
label1 = filter_labels[count:i]
|
473 |
key1 = h[i - 1]
|
|
|
479 |
for i in range(len(label1)):
|
480 |
vote = label1[i]
|
481 |
classcount[vote] = classcount.get(vote, 0) + 1
|
482 |
+
|
483 |
sortedclass = sorted(classcount.items(), key=lambda x: (x[1]), reverse=True)
|
|
|
484 |
point_index = same_points_list[key1]
|
485 |
for j in range(len(point_index)):
|
486 |
upsample_label.append(sortedclass[0][0])
|
|
|
488 |
upsample_point.append(point_cloud[index])
|
489 |
upsample_index.append(index)
|
490 |
|
|
|
491 |
classcount = {}
|
492 |
for i in range(len(label2)):
|
493 |
vote = label2[i]
|
494 |
classcount[vote] = classcount.get(vote, 0) + 1
|
495 |
+
|
496 |
sortedclass = sorted(classcount.items(), key=lambda x: (x[1]), reverse=True)
|
|
|
497 |
point_index = same_points_list[key2]
|
498 |
for j in range(len(point_index)):
|
499 |
upsample_label.append(sortedclass[0][0])
|
|
|
508 |
for i in range(len(label)):
|
509 |
vote = label[i]
|
510 |
classcount[vote] = classcount.get(vote, 0) + 1
|
511 |
+
|
512 |
sortedclass = sorted(classcount.items(), key=lambda x: (x[1]), reverse=True)
|
|
|
513 |
point_index = same_points_list[key] # h对应的point index列表
|
514 |
for j in range(len(point_index)):
|
515 |
upsample_label.append(sortedclass[0][0])
|
516 |
index = point_index[j]
|
517 |
upsample_point.append(point_cloud[index])
|
518 |
upsample_index.append(index)
|
|
|
519 |
|
520 |
+
# Restore the original order of index
|
|
|
|
|
|
|
|
|
521 |
upsample_index = np.array(upsample_index)
|
522 |
+
upsample_index_indice = np.argsort(upsample_index) # Extract the index and return the index of the elements in h sorted from small to large.
|
523 |
upsample_index_sorted = upsample_index[upsample_index_indice]
|
524 |
|
525 |
upsample_point = np.array(upsample_point)
|
526 |
upsample_label = np.array(upsample_label)
|
527 |
+
|
528 |
+
# Restore the original order of points and labels
|
529 |
upsample_point_sorted = upsample_point[upsample_index_indice]
|
530 |
upsample_label_sorted = upsample_label[upsample_index_indice]
|
531 |
|
532 |
return upsample_point_sorted, upsample_label_sorted
|
533 |
|
534 |
+
# Upsampling using knn algorithm
|
|
|
535 |
def KNN_sklearn_Load_data(voxel_points, center_points, labels):
|
536 |
+
# Build model
|
|
|
|
|
537 |
model = neighbors.KNeighborsClassifier(n_neighbors=3)
|
538 |
model.fit(center_points, labels)
|
539 |
prediction = model.predict(voxel_points.reshape(1, -1))
|
|
|
|
|
540 |
|
541 |
+
return prediction[0]
|
542 |
|
543 |
+
# Loading points for knn upsampling
|
544 |
def Load_data(voxel_points, center_points, labels):
|
545 |
meshtopoints_labels = []
|
|
|
546 |
for i in range(0, voxel_points.shape[0]):
|
547 |
meshtopoints_labels.append(KNN_sklearn_Load_data(voxel_points[i], center_points, labels))
|
548 |
+
|
549 |
return np.array(meshtopoints_labels)
|
550 |
|
551 |
+
# Upsample triangular mesh data back to original point cloud data
|
552 |
def mesh_to_points_main(jaw, pcd_points, center_points, labels):
|
553 |
points = pcd_points.copy()
|
554 |
+
|
555 |
+
# Downsampling
|
556 |
voxel_points, same_points_list = voxel_filter(points, 0.6)
|
557 |
|
558 |
after_labels = Load_data(voxel_points, center_points, labels)
|
|
|
562 |
new_pcd = o3d.geometry.PointCloud()
|
563 |
new_pcd.points = o3d.utility.Vector3dVector(upsample_point)
|
564 |
instances_labels = upsample_label.copy()
|
565 |
+
|
566 |
+
# Reclassify the label of the upper and lower jaws
|
567 |
for i in stqdm(range(0, upsample_label.shape[0])):
|
568 |
if jaw == 'upper':
|
569 |
if (upsample_label[i] >= 1) and (upsample_label[i] <= 8):
|
|
|
575 |
upsample_label[i] = upsample_label[i] + 30
|
576 |
elif (upsample_label[i] >= 9) and (upsample_label[i] <= 16):
|
577 |
upsample_label[i] = upsample_label[i] + 32
|
578 |
+
|
579 |
remove_outlier_main(jaw, pcd_points, upsample_label, instances_labels)
|
580 |
|
581 |
|
582 |
+
# Convert raw point cloud data to triangular mesh
|
583 |
def mesh_grid(pcd_points):
|
584 |
new_pcd,_ = voxel_filter(pcd_points, 0.6)
|
585 |
+
# pcd needs to have a normal vector
|
586 |
|
587 |
# estimate radius for rolling ball
|
588 |
pcd_new = o3d.geometry.PointCloud()
|
|
|
594 |
mesh = o3d.geometry.TriangleMesh.create_from_point_cloud_ball_pivoting(
|
595 |
pcd_new,
|
596 |
o3d.utility.DoubleVector([radius, radius * 2]))
|
|
|
597 |
|
598 |
return mesh
|
599 |
|
600 |
+
# Read the contents of obj file
|
|
|
601 |
def read_obj(obj_path):
|
602 |
jaw = None
|
603 |
with open(obj_path) as file:
|
|
|
619 |
|
620 |
points = np.array(points)
|
621 |
faces = np.array(faces)
|
|
|
622 |
if jaw is None:
|
623 |
raise ValueError("Jaw type not found in OBJ file")
|
624 |
|
625 |
return points, faces, jaw
|
626 |
|
627 |
+
# Convert obj file to pcd file
|
|
|
628 |
def obj2pcd(obj_path):
|
629 |
if os.path.exists(obj_path):
|
630 |
print('yes')
|
|
|
636 |
pcd_list.append(new_line.split())
|
637 |
|
638 |
pcd_points = np.array(pcd_list).astype(np.float64)
|
|
|
639 |
|
640 |
+
return pcd_points, jaw
|
641 |
|
642 |
+
# Main function for segment
|
643 |
def segmentation_main(obj_path):
|
644 |
upsampling_method = 'KNN'
|
645 |
|
646 |
+
model_path = 'model.tar'
|
647 |
num_classes = 17
|
648 |
num_channels = 15
|
649 |
|
|
|
713 |
nmeans = normals.mean(axis=0)
|
714 |
nstds = normals.std(axis=0)
|
715 |
|
716 |
+
# normalization
|
717 |
for i in range(3):
|
718 |
cells[:, i] = (cells[:, i] - means[i]) / stds[i] # point 1
|
719 |
cells[:, i + 3] = (cells[:, i + 3] - means[i]) / stds[i] # point 2
|
|
|
721 |
barycenters[:, i] = (barycenters[:, i] - mins[i]) / (maxs[i] - mins[i])
|
722 |
normals[:, i] = (normals[:, i] - nmeans[i]) / nstds[i]
|
723 |
|
724 |
+
# concatenate
|
725 |
X = np.column_stack((cells, barycenters, normals))
|
726 |
|
727 |
# computing A_S and A_L
|
|
|
772 |
if i_node < i_nei:
|
773 |
cos_theta = np.dot(normals[i_node, 0:3], normals[i_nei, 0:3]) / np.linalg.norm(
|
774 |
normals[i_node, 0:3]) / np.linalg.norm(normals[i_nei, 0:3])
|
775 |
+
|
776 |
if cos_theta >= 1.0:
|
777 |
cos_theta = 0.9999
|
778 |
theta = np.arccos(cos_theta)
|
|
|
785 |
edges = np.concatenate(
|
786 |
(edges, np.array([i_node, i_nei, -beta * np.log10(theta / np.pi) * phi]).reshape(1, 3)),
|
787 |
axis=0)
|
788 |
+
|
789 |
edges = np.delete(edges, 0, 0)
|
790 |
edges[:, 2] *= lambda_c * round_factor
|
791 |
edges = edges.astype(np.int32)
|
|
|
893 |
# Create a pyvista plotter
|
894 |
plotter = pv.Plotter()
|
895 |
|
896 |
+
cmap = plt.cm.get_cmap('jet', 27) # Using a colormap with sufficient distinct colors
|
897 |
|
898 |
+
colors = cmap(np.linspace(0, 1, 27)) # Generate colors
|
899 |
|
900 |
# Convert colors to a format acceptable by PyVista
|
901 |
colormap = mcolors.ListedColormap(colors)
|
|
|
910 |
with st.expander("Ground Truth - scroll for zoom", expanded=False):
|
911 |
stpyvista(plotter)
|
912 |
|
|
|
|
|
913 |
elif inputs == "Upload Scan":
|
914 |
file = st.file_uploader("Please upload an OBJ Object file", type=["OBJ"])
|
915 |
st.markdown("Expected time per prediction: 7-10 min.")
|
|
|
917 |
# save the uploaded file to disk
|
918 |
with open("file.obj", "wb") as buffer:
|
919 |
shutil.copyfileobj(file, buffer)
|
920 |
+
|
921 |
obj_path = "file.obj"
|
922 |
|
923 |
mesh = pv.read(obj_path)
|
|
|
935 |
if segment:
|
936 |
segmentation_main(obj_path)
|
937 |
|
|
|
|
|
|
|
|
|
938 |
if __name__ == "__main__":
|
939 |
app = Segment()
|