forked from Rhett98/dynamic_slam
-
Notifications
You must be signed in to change notification settings - Fork 0
/
dylo_model.py
555 lines (439 loc) · 36 KB
/
dylo_model.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
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
# -*- coding:UTF-8 -*-
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
import time
import numpy as np
from conv_util import PointNetSaModule, cost_volume, set_upconv_module, FlowPredictor, Conv1d
from model_utils import ProjectPCimg2SphericalRing, PreProcess, mat2euler, euler2quat, \
softmax_valid, quat2mat, inv_q, mul_q_point, mul_point_q
scale = 1.0
def get_selected_idx(batch_size, out_H: int, out_W: int, stride_H: int, stride_W: int):
"""According to given stride and output size, return the corresponding selected points
Args:
array (tf.Tensor): [any array with shape (B, H, W, 3)]
stride_H (int): [stride in height]
stride_W (int): [stride in width]
out_H (int): [height of output array]
out_W (int): [width of output array]
Returns:
[tf.Tensor]: [shape (B, outh, outw, 3) indices]
"""
select_h_idx = torch.arange(0, out_H * stride_H, stride_H)
select_w_idx = torch.arange(0, out_W * stride_W, stride_W)
height_indices = (torch.reshape(select_h_idx, (1, -1, 1))).expand(batch_size, out_H, out_W) # b out_H out_W
width_indices = (torch.reshape(select_w_idx, (1, 1, -1))).expand(batch_size, out_H, out_W) # b out_H out_W
padding_indices = torch.reshape(torch.arange(batch_size), (-1, 1, 1)).expand(batch_size, out_H, out_W) # b out_H out_W
return padding_indices, height_indices, width_indices
class dylo_model(nn.Module):
def __init__(self, args, batch_size, H_input, W_input, is_training, bn_decay=None):
super(dylo_model, self).__init__()
##### initialize the parameters (distance & stride ) ######
self.H_input = H_input; self.W_input = W_input
self.Down_conv_dis = [0.75, 3.0, 6.0, 12.0]
self.Up_conv_dis = [3.0, 6.0, 9.0]
self.Cost_volume_dis = [1.0, 2.0, 4.5]
self.stride_H_list = [4, 2, 2, 1]
self.stride_W_list = [8, 2, 2, 2]
self.out_H_list = [math.ceil(self.H_input / self.stride_H_list[0])] #
self.out_W_list = [math.ceil(self.W_input / self.stride_W_list[0])] #self.W_input
for i in range(1, 4):
self.out_H_list.append(math.ceil(self.out_H_list[i - 1] / self.stride_H_list[i])) ##(16,8,4,4)
self.out_W_list.append(math.ceil(self.out_W_list[i - 1] / self.stride_W_list[i])) ##(57,29,15,8) # generate the output shape list
self.training = is_training
self.w_x = nn.Parameter(torch.tensor([0.0]), requires_grad=True)
self.w_q = nn.Parameter(torch.tensor([-2.5]), requires_grad=True)
################################
# Stride-based Sampling #
################################
self.layer0 = PointNetSaModule(batch_size = batch_size, K_sample = 32, kernel_size = [9, 15], H = self.out_H_list[0], W = self.out_W_list[0], \
stride_H = self.stride_H_list[0], stride_W = self.stride_W_list[0], distance = self.Down_conv_dis[0], in_channels = 3,
mlp = [8, 8, 16], is_training = self.training,
bn_decay = bn_decay)
self.layer1 = PointNetSaModule(batch_size = batch_size, K_sample = 32, kernel_size = [7, 11], H = self.out_H_list[1], W = self.out_W_list[1], \
stride_H = self.stride_H_list[1], stride_W = self.stride_W_list[1], distance = self.Down_conv_dis[1],
in_channels = 16,
mlp=[16, 16, 32], is_training=self.training,
bn_decay = bn_decay)
self.layer2 = PointNetSaModule(batch_size = batch_size, K_sample = 16, kernel_size = [5, 9], H = self.out_H_list[2], W = self.out_W_list[2], \
stride_H = self.stride_H_list[2], stride_W = self.stride_W_list[2], distance = self.Down_conv_dis[2],
in_channels=32,
mlp=[32, 32, 64], is_training=self.training,
bn_decay=bn_decay)
self.layer3 = PointNetSaModule(batch_size = batch_size, K_sample = 16, kernel_size = [5, 9], H = self.out_H_list[3], W = self.out_W_list[3], \
stride_H = self.stride_H_list[3], stride_W = self.stride_W_list[3], distance = self.Down_conv_dis[3],
in_channels=64,
mlp=[64, 64, 128], is_training=self.training,
bn_decay=bn_decay)
self.laye3_1 = PointNetSaModule(batch_size = batch_size, K_sample = 16, kernel_size = [5, 9], H = self.out_H_list[3], W = self.out_W_list[3], \
stride_H = self.stride_H_list[3], stride_W = self.stride_W_list[3], distance = self.Down_conv_dis[3],
in_channels=64,
mlp=[128, 64, 64], is_training=self.training,
bn_decay=bn_decay)
#############################
# Cost volume #
#############################
self.cost_volume1 = cost_volume(batch_size=batch_size, kernel_size1=[3, 5], kernel_size2=[5, 35], nsample=4,
nsample_q=32, \
H=self.out_H_list[2], W=self.out_W_list[2], \
stride_H=1, stride_W=1, distance=self.Cost_volume_dis[2],
in_channels=[64, 64],
mlp1=[128, 64, 64], mlp2=[128, 64], is_training=self.training,
bn_decay=bn_decay,
bn=True, pooling='max', knn=True, corr_func='concat')
self.cost_volume2 = cost_volume(batch_size=batch_size, kernel_size1=[3, 5], kernel_size2=[5, 15], nsample=4,
nsample_q=6, \
H=self.out_H_list[2], W=self.out_W_list[2], \
stride_H=1, stride_W=1, distance=self.Cost_volume_dis[2],
in_channels=[64, 64],
mlp1=[128, 64, 64], mlp2=[128, 64], is_training=self.training,
bn_decay=bn_decay,
bn=True, pooling='max', knn=True, corr_func='concat')
self.cost_volume3 = cost_volume(batch_size=batch_size, kernel_size1=[3, 5], kernel_size2=[7, 25], nsample=4,
nsample_q=6, \
H=self.out_H_list[1], W=self.out_W_list[1], \
stride_H=1, stride_W=1, distance=self.Cost_volume_dis[1],
in_channels=[32, 32],
mlp1=[128, 64, 64], mlp2=[128, 64], is_training=self.training,
bn_decay=bn_decay,
bn=True, pooling='max', knn=True, corr_func='concat')
self.cost_volume4 = cost_volume(batch_size=batch_size, kernel_size1=[3, 5], kernel_size2=[11, 41], nsample=4,
nsample_q=6, \
H=self.out_H_list[0], W=self.out_W_list[0], \
stride_H=1, stride_W=1, distance=self.Cost_volume_dis[0],
in_channels=[16, 16],
mlp1=[128, 64, 64], mlp2=[128, 64], is_training=self.training,
bn_decay=bn_decay,
bn=True, pooling='max', knn=True, corr_func='concat')
###############################
# MLP to predict flow#
###############################
self.flow_predictor0 = FlowPredictor(in_channels=64 * 3, mlp=[128, 64], is_training=self.training,
bn_decay=bn_decay)
self.flow_predictor1_predict = FlowPredictor(in_channels=64 * 3, mlp=[128, 64], is_training=self.training,
bn_decay=bn_decay)
self.flow_predictor1_w = FlowPredictor(in_channels=64 * 3, mlp=[128, 64], is_training=self.training,
bn_decay=bn_decay)
self.flow_predictor2_predict = FlowPredictor(in_channels=64 * 2 + 32, mlp=[128, 64], is_training=self.training,
bn_decay=bn_decay)
self.flow_predictor2_w = FlowPredictor(in_channels=64 * 2 + 32, mlp=[128, 64], is_training=self.training,
bn_decay=bn_decay)
self.flow_predictor3_predict = FlowPredictor(in_channels=64 * 2 + 16, mlp=[128, 64], is_training=self.training,
bn_decay=bn_decay)
self.flow_predictor3_w = FlowPredictor(in_channels=64 * 2 + 16, mlp=[128, 64], is_training=self.training,
bn_decay=bn_decay)
###################################
# Up-sampling layers #
###################################
self.set_upconv1_w_upsample = set_upconv_module(batch_size=batch_size, kernel_size=[7, 15],
H=self.out_H_list[2], W=self.out_W_list[2],
stride_H=self.stride_H_list[-1],
stride_W=self.stride_W_list[-1],
nsample=8, distance=self.Up_conv_dis[2],
in_channels=[64, 64],
mlp=[128, 64], mlp2=[64], is_training=self.training,
bn_decay=bn_decay, knn=True)
self.set_upconv1_upsample = set_upconv_module(batch_size=batch_size, kernel_size=[7, 15],
H=self.out_H_list[2], W=self.out_W_list[2],
stride_H=self.stride_H_list[-1], stride_W=self.stride_W_list[-1],
nsample=8, distance=self.Up_conv_dis[2],
in_channels=[64, 64],
mlp=[128, 64], mlp2=[64], is_training=self.training,
bn_decay=bn_decay, knn=True)
self.set_upconv2_w_upsample = set_upconv_module(batch_size=batch_size, kernel_size=[7, 15],
H=self.out_H_list[1], W=self.out_W_list[1],
stride_H=self.stride_H_list[-2],
stride_W=self.stride_W_list[-2], \
nsample=8, distance=self.Up_conv_dis[1],
in_channels=[32, 64],
mlp=[128, 64], mlp2=[64], is_training=self.training,
bn_decay=bn_decay, knn=True)
self.set_upconv2_upsample = set_upconv_module(batch_size=batch_size, kernel_size=[7, 15],
H=self.out_H_list[1], W=self.out_W_list[1],
stride_H=self.stride_H_list[-2], stride_W=self.stride_W_list[-2], \
nsample=8, distance=self.Up_conv_dis[1],
in_channels=[32, 64],
mlp=[128, 64], mlp2=[64], is_training=self.training,
bn_decay=bn_decay, knn=True)
self.set_upconv3_w_upsample = set_upconv_module(batch_size=batch_size, kernel_size=[7, 15],
H=self.out_H_list[0], W=self.out_W_list[0],
stride_H=self.stride_H_list[-3],
stride_W=self.stride_W_list[-3], \
nsample=8, distance=self.Up_conv_dis[0],
in_channels=[16, 64],
mlp=[128, 64], mlp2=[64], is_training=self.training,
bn_decay=bn_decay, knn=True)
self.set_upconv3_upsample = set_upconv_module(batch_size=batch_size, kernel_size=[7, 15],
H=self.out_H_list[0], W=self.out_W_list[0],
stride_H=self.stride_H_list[-3], stride_W=self.stride_W_list[-3], \
nsample=8, distance=self.Up_conv_dis[0],
in_channels=[16, 64],
mlp=[128, 64], mlp2=[64], is_training=self.training,
bn_decay=bn_decay, knn=True)
###################################################
# Conv layers to regress pose #
###################################################
self.conv1_l3 = Conv1d(256, 4, use_activation=False)
self.conv1_l2 = Conv1d(256, 4, use_activation=False)
self.conv1_l1 = Conv1d(256, 4, use_activation=False)
self.conv1_l0 = Conv1d(256, 4, use_activation=False)
self.conv2_l3 = Conv1d(256, 3, use_activation=False)
self.conv2_l2 = Conv1d(256, 3, use_activation=False)
self.conv2_l1 = Conv1d(256, 3, use_activation=False)
self.conv2_l0 = Conv1d(256, 3, use_activation=False)
self.conv3_l3 = Conv1d(64, 256, use_activation=False)
self.conv3_l2 = Conv1d(64, 256, use_activation=False)
self.conv3_l1 = Conv1d(64, 256, use_activation=False)
self.conv3_l0 = Conv1d(64, 256, use_activation=False)
def forward(self, input_xyz_f1, input_xyz_f2, T_gt, T_trans, T_trans_inv):
start_train = time.time()
batch_size = len(input_xyz_f1)
torch.cuda.synchronize()
start_time = time.time()
aug_frame = np.random.choice([1, 2], size = batch_size, replace = True) # random choose aug frame 1 or 2
input_xyz_aug_f1, input_xyz_aug_f2, q_gt, t_gt = PreProcess(input_xyz_f1, input_xyz_f2, T_gt, T_trans, T_trans_inv, aug_frame)
####proj func(2D3D)
# 点云柱面投影，返回image和对应的(B,n,3)的点云
input_xyz_aug_proj_f1, mask_xyz_f1 = ProjectPCimg2SphericalRing(input_xyz_aug_f1, None, self.H_input, self.W_input) ## proj func
input_xyz_aug_proj_f2, mask_xyz_f2 = ProjectPCimg2SphericalRing(input_xyz_aug_f2, None, self.H_input, self.W_input)
# 对点云image进行降采样，该步得到降采样的索引
self.l0_b_idx, self.l0_h_idx, self.l0_w_idx = get_selected_idx(batch_size, self.out_H_list[0],
self.out_W_list[0], self.stride_H_list[0],
self.stride_W_list[0])
self.l1_b_idx, self.l1_h_idx, self.l1_w_idx = get_selected_idx(batch_size, self.out_H_list[1],
self.out_W_list[1], self.stride_H_list[1],
self.stride_W_list[1])
self.l2_b_idx, self.l2_h_idx, self.l2_w_idx = get_selected_idx(batch_size, self.out_H_list[2],
self.out_W_list[2], self.stride_H_list[2],
self.stride_W_list[2])
self.l3_b_idx, self.l3_h_idx, self.l3_w_idx = get_selected_idx(batch_size, self.out_H_list[3],
self.out_W_list[3], self.stride_H_list[3],
self.stride_W_list[3])
###########################
# Kernel center #
###########################
# 获取降采样后的点云
#### the l0 select bn3 xyz
l0_xyz_proj_f1 = input_xyz_aug_proj_f1[self.l0_b_idx.cuda().long(), self.l0_h_idx.cuda().long(), self.l0_w_idx.cuda().long(), :] #### PC1，PC2
l0_xyz_proj_f2 = input_xyz_aug_proj_f2[self.l0_b_idx.cuda().long(), self.l0_h_idx.cuda().long(), self.l0_w_idx.cuda().long(), :]
#### the l1 select bn3 xyz
l1_xyz_proj_f1 = l0_xyz_proj_f1[self.l1_b_idx.cuda().long(), self.l1_h_idx.cuda().long(), self.l1_w_idx.cuda().long(), :]
l1_xyz_proj_f2 = l0_xyz_proj_f2[self.l1_b_idx.cuda().long(), self.l1_h_idx.cuda().long(), self.l1_w_idx.cuda().long(), :]
#### the l2 select bn3 xyz
l2_xyz_proj_f1 = l1_xyz_proj_f1[self.l2_b_idx.cuda().long(), self.l2_h_idx.cuda().long(), self.l2_w_idx.cuda().long(), :]
l2_xyz_proj_f2 = l1_xyz_proj_f2[self.l2_b_idx.cuda().long(), self.l2_h_idx.cuda().long(), self.l2_w_idx.cuda().long(), :]
#### the l3 select bn3 xyz
l3_xyz_proj_f1 = l2_xyz_proj_f1[self.l3_b_idx.cuda().long(), self.l3_h_idx.cuda().long(), self.l3_w_idx.cuda().long(), :]
l3_xyz_proj_f2 = l2_xyz_proj_f2[self.l3_b_idx.cuda().long(), self.l3_h_idx.cuda().long(), self.l3_w_idx.cuda().long(), :]
###########################
# Binary masks #
###########################
#### the l0 select bn1 mask
l0_mask_f1 = mask_xyz_f1[self.l0_b_idx.cuda().long(), self.l0_h_idx.cuda().long(), self.l0_w_idx.cuda().long(), :] #### PC1，PC2
l0_mask_f2 = mask_xyz_f2[self.l0_b_idx.cuda().long(), self.l0_h_idx.cuda().long(), self.l0_w_idx.cuda().long(), :]
#### the l1 select bn1 mask
l1_mask_f1 = l0_mask_f1[self.l1_b_idx.cuda().long(), self.l1_h_idx.cuda().long(), self.l1_w_idx.cuda().long(), :]
l1_mask_f2 = l0_mask_f2[self.l1_b_idx.cuda().long(), self.l1_h_idx.cuda().long(), self.l1_w_idx.cuda().long(), :]
#### the l2 select bn1 mask
l2_mask_f1 = l1_mask_f1[self.l2_b_idx.cuda().long(), self.l2_h_idx.cuda().long(), self.l2_w_idx.cuda().long(), :]
l2_mask_f2 = l1_mask_f2[self.l2_b_idx.cuda().long(), self.l2_h_idx.cuda().long(), self.l2_w_idx.cuda().long(), :]
#### the l3 select bn1 mask
l3_mask_f1 = l2_mask_f1[self.l3_b_idx.cuda().long(), self.l3_h_idx.cuda().long(), self.l3_w_idx.cuda().long(), :]
l3_mask_f2 = l2_mask_f2[self.l3_b_idx.cuda().long(), self.l3_h_idx.cuda().long(), self.l3_w_idx.cuda().long(), :]
###set conv
# 特征提取
set_conv_start = time.time()
input_points_f1 = torch.zeros_like(input_xyz_aug_proj_f1)
input_points_f2 = torch.zeros_like(input_xyz_aug_proj_f2)
# Flame 1
l0_points_f1, l0_points_proj_f1 = self.layer0(input_xyz_aug_proj_f1, input_points_f1, l0_xyz_proj_f1) ###l0层的F1
l0_points_proj_f1 = torch.reshape(l0_points_f1, (batch_size, self.out_H_list[0], self.out_W_list[0], -1))
l1_points_f1, l1_points_proj_f1 = self.layer1(l0_xyz_proj_f1, l0_points_proj_f1, l1_xyz_proj_f1)
l1_points_proj_f1 = torch.reshape(l1_points_f1, (batch_size, self.out_H_list[1], self.out_W_list[1], -1))
l2_points_f1, l2_points_proj_f1 = self.layer2(l1_xyz_proj_f1, l1_points_proj_f1, l2_xyz_proj_f1)
l2_points_proj_f1 = torch.reshape(l2_points_f1, (batch_size, self.out_H_list[2], self.out_W_list[2], -1))
##### Flame 2
l0_points_f2, l0_points_proj_f2 = self.layer0(input_xyz_aug_proj_f2, input_points_f2, l0_xyz_proj_f2)
l0_points_proj_f2 = torch.reshape(l0_points_f2, (batch_size, self.out_H_list[0], self.out_W_list[0], -1))
l1_points_f2, l1_points_proj_f2 = self.layer1(l0_xyz_proj_f2, l0_points_proj_f2, l1_xyz_proj_f2)
l1_points_proj_f2 = torch.reshape(l1_points_f2, (batch_size, self.out_H_list[1], self.out_W_list[1], -1))
l2_points_f2, l2_points_proj_f2 = self.layer2(l1_xyz_proj_f2, l1_points_proj_f2, l2_xyz_proj_f2)
l2_points_proj_f2 = torch.reshape(l2_points_f2, (batch_size, self.out_H_list[2], self.out_W_list[2], -1))
###cost volume
l2_cost_volume_origin = self.cost_volume1(l2_xyz_proj_f1, l2_xyz_proj_f2, l2_points_proj_f1, l2_points_proj_f2)
l2_cost_volume_origin_proj = torch.reshape(l2_cost_volume_origin, [batch_size, self.out_H_list[2], self.out_W_list[2], -1]) ##E2
# Layer 3 ##################
l3_points_f1, l3_points_proj_f1 = self.layer3(l2_xyz_proj_f1, l2_points_proj_f1, l3_xyz_proj_f1)
l3_points_f2, l3_points_proj_f2 = self.layer3(l2_xyz_proj_f2, l2_points_proj_f2, l3_xyz_proj_f2)
l3_cost_volume, l3_cost_volume_proj = self.laye3_1(l2_xyz_proj_f1, l2_cost_volume_origin_proj, l3_xyz_proj_f1) ##E3
# print("l3_points_f1:",l3_points_f1.shape)
###l3 cost_volume
l3_cost_volume_w = self.flow_predictor0(l3_points_f1, None, l3_cost_volume) ##E3和F3_f1融合
# print("l3_cost_volume_w:", l3_cost_volume_w.shape)
l3_cost_volume_w_proj = torch.reshape(l3_cost_volume_w, [batch_size, self.out_H_list[3], self.out_W_list[3], -1])
l3_xyz_f1 = torch.reshape(l3_xyz_proj_f1, [batch_size, -1, 3])
mask_l3 = torch.any(l3_xyz_f1 != 0, dim=-1)
l3_points_f1_new = softmax_valid(feature_bnc=l3_cost_volume, weight_bnc=l3_cost_volume_w, mask_valid=mask_l3) # B 1 C
l3_points_f1_new_big = self.conv3_l3(l3_points_f1_new)
l3_points_f1_new_q = F.dropout(l3_points_f1_new_big, p=0.5, training=self.training)
l3_points_f1_new_t = F.dropout(l3_points_f1_new_big, p=0.5, training=self.training)
l3_q_coarse = self.conv1_l3(l3_points_f1_new_q)
l3_q_coarse = l3_q_coarse / (torch.sqrt(torch.sum(l3_q_coarse * l3_q_coarse, dim=-1, keepdim=True) + 1e-10) + 1e-10)
l3_t_coarse = self.conv2_l3(l3_points_f1_new_t)
l3_q = torch.squeeze(l3_q_coarse, dim=1)
l3_t = torch.squeeze(l3_t_coarse, dim=1)
################ layer 2 #################
l2_q_coarse = torch.reshape(l3_q, [batch_size, 1, -1])
l2_t_coarse = torch.reshape(l3_t, [batch_size, 1, -1])
l2_q_inv = inv_q(l2_q_coarse, batch_size)
### warp layer2 pose
l2_xyz_f1 = torch.reshape(l2_xyz_proj_f1, [batch_size, -1, 3])
l2_xyz_bnc_q = torch.cat([torch.zeros([batch_size, self.out_H_list[2] * self.out_W_list[2], 1]).cuda(), l2_xyz_f1], dim=-1)
l2_flow_warped = mul_q_point(l2_q_coarse, l2_xyz_bnc_q, batch_size)
l2_flow_warped = torch.index_select(mul_point_q(l2_flow_warped, l2_q_inv, batch_size), 2, torch.LongTensor(range(1, 4)).cuda()) + l2_t_coarse
l2_mask = torch.any(l2_xyz_f1 != 0, dim=-1, keepdim=True).to(torch.float32)
l2_flow_warped = l2_flow_warped * l2_mask
### re-project
l2_xyz_warp_proj_f1, l2_points_warp_proj_f1 = ProjectPCimg2SphericalRing(l2_flow_warped, l2_points_f1, self.out_H_list[2], self.out_W_list[2]) #
l2_xyz_warp_f1 = torch.reshape(l2_xyz_warp_proj_f1, [batch_size, -1, 3])
l2_points_warp_f1 = torch.reshape(l2_points_warp_proj_f1, [batch_size, self.out_H_list[2] * self.out_W_list[2], -1])
l2_mask_warped = torch.any(l2_xyz_warp_f1 != 0, dim=-1, keepdim=False)
l2_mask_warped_proj = torch.reshape(l2_mask_warped, [batch_size, self.out_H_list[2], self.out_W_list[2], -1])
# print(l2_points_warp_proj_f1.shape)
# get the cost volume of warped layer2 flow and the points of frame2
l2_points_warp_proj_f1 = torch.reshape(l2_points_warp_f1, [batch_size, self.out_H_list[2], self.out_W_list[2], -1])
l2_points_proj_f2 = torch.reshape(l2_points_f2, [batch_size, self.out_H_list[2], self.out_W_list[2], -1])
l2_cost_volume = self.cost_volume2(l2_xyz_warp_proj_f1, l2_xyz_proj_f2, l2_points_warp_proj_f1, l2_points_proj_f2) ##RE
l2_cost_volume_w_upsample = self.set_upconv1_w_upsample(l2_xyz_warp_proj_f1, l3_xyz_proj_f1, l2_points_warp_proj_f1, l3_cost_volume_w_proj) ###CE
l2_cost_volume_upsample = self.set_upconv1_upsample(l2_xyz_warp_proj_f1, l3_xyz_proj_f1, l2_points_warp_proj_f1, l3_cost_volume_proj)
# print("l2_points_warp_f1:",l2_points_warp_f1.shape)
l2_cost_volume_predict = self.flow_predictor1_predict(l2_points_warp_f1, l2_cost_volume_upsample, l2_cost_volume) ##E
l2_cost_volume_w = self.flow_predictor1_w(l2_points_warp_f1, l2_cost_volume_w_upsample, l2_cost_volume_predict)
# print("l2_cost_volume_w:", l2_cost_volume_w.shape)
l2_cost_volume_proj = torch.reshape(l2_cost_volume_predict, [batch_size, self.out_H_list[2], self.out_W_list[2], -1])
l2_cost_volume_w_proj = torch.reshape(l2_cost_volume_w, [batch_size, self.out_H_list[2], self.out_W_list[2], -1])
l2_cost_volume_sum = softmax_valid(feature_bnc=l2_cost_volume_predict, weight_bnc=l2_cost_volume_w, mask_valid=l2_mask_warped) # embedding mask B 1 C
l2_points_f1_new_big = self.conv3_l2(l2_cost_volume_sum)
l2_points_f1_new_q = F.dropout(l2_points_f1_new_big, p=0.5, training=self.training)
l2_points_f1_new_t = F.dropout(l2_points_f1_new_big, p=0.5, training=self.training)
l2_q_det = self.conv1_l2(l2_points_f1_new_q)
l2_q_det = l2_q_det / (torch.sqrt(torch.sum(l2_q_det * l2_q_det, dim=-1, keepdim=True) + 1e-10) + 1e-10)
l2_q_det_inv = inv_q(l2_q_det, batch_size)
l2_t_det = self.conv2_l2(l2_points_f1_new_t)
l2_t_coarse_trans = torch.cat([torch.zeros([batch_size, 1, 1]).cuda(), l2_t_coarse], dim=-1)
l2_t_coarse_trans = mul_q_point(l2_q_det, l2_t_coarse_trans, batch_size)
l2_t_coarse_trans = torch.index_select(mul_point_q(l2_t_coarse_trans, l2_q_det_inv, batch_size), 2,
torch.LongTensor(range(1, 4)).cuda())
l2_q = torch.squeeze(mul_point_q(l2_q_det, l2_q_coarse, batch_size), dim=1)
l2_t = torch.squeeze(l2_t_coarse_trans + l2_t_det, dim=1)
############# layer1
start_l1_refine = time.time()
l1_q_coarse = torch.reshape(l2_q, [batch_size, 1, -1])
l1_t_coarse = torch.reshape(l2_t, [batch_size, 1, -1])
l1_q_inv = inv_q(l1_q_coarse, batch_size)
############# warp layer2 pose
l1_xyz_f1 = torch.reshape(l1_xyz_proj_f1, [batch_size, -1, 3])
l1_xyz_bnc_q = torch.cat([torch.zeros([batch_size, self.out_H_list[1] * self.out_W_list[1], 1]).cuda(), l1_xyz_f1], dim=-1)
l1_flow_warped = mul_q_point(l1_q_coarse, l1_xyz_bnc_q, batch_size)
l1_flow_warped = torch.index_select(mul_point_q(l1_flow_warped, l1_q_inv, batch_size), 2, torch.LongTensor(range(1, 4)).cuda()) + l1_t_coarse
l1_mask = torch.any(l1_xyz_f1 != 0, dim=-1, keepdim=True).to(torch.float32)
l1_flow_warped = l1_flow_warped * l1_mask
########## re-project
l1_xyz_warp_proj_f1, l1_points_warp_proj_f1 = ProjectPCimg2SphericalRing(l1_flow_warped, l1_points_f1, self.out_H_list[1], self.out_W_list[1]) #
l1_xyz_warp_f1 = torch.reshape(l1_xyz_warp_proj_f1, [batch_size, -1, 3])
l1_points_warp_f1 = torch.reshape(l1_points_warp_proj_f1, [batch_size, self.out_H_list[1] * self.out_W_list[1], -1])
l1_mask_warped = torch.any(l1_xyz_warp_f1 != 0, dim=-1, keepdim=False)
l1_mask_warped_proj = torch.reshape(l1_mask_warped, [batch_size, self.out_H_list[1], self.out_W_list[1], -1])
# get the cost volume of warped layer1 flow and the points of frame2
l1_points_warp_proj_f1 = torch.reshape(l1_points_warp_f1, [batch_size, self.out_H_list[1], self.out_W_list[1], -1])
l1_points_proj_f2 = torch.reshape(l1_points_f2, [batch_size, self.out_H_list[1], self.out_W_list[1], -1])
l1_cost_volume = self.cost_volume3(l1_xyz_warp_proj_f1, l1_xyz_proj_f2, l1_points_warp_proj_f1, l1_points_proj_f2)
l1_cost_volume_w_upsample = self.set_upconv2_w_upsample(l1_xyz_warp_proj_f1, l2_xyz_warp_proj_f1, l1_points_warp_proj_f1, l2_cost_volume_w_proj)
l1_cost_volume_upsample = self.set_upconv2_upsample(l1_xyz_warp_proj_f1, l2_xyz_warp_proj_f1, l1_points_warp_proj_f1, l2_cost_volume_proj)
l1_cost_volume_predict = self.flow_predictor2_predict(l1_points_warp_f1, l1_cost_volume_upsample, l1_cost_volume)
l1_cost_volume_w = self.flow_predictor2_w(l1_points_warp_f1, l1_cost_volume_w_upsample, l1_cost_volume_predict)
l1_cost_volume_proj = torch.reshape(l1_cost_volume_predict, [batch_size, self.out_H_list[1], self.out_W_list[1], -1])
l1_cost_volume_w_proj = torch.reshape(l1_cost_volume_w, [batch_size, self.out_H_list[1], self.out_W_list[1], -1])
l1_cost_volume_sum = softmax_valid(feature_bnc=l1_cost_volume_predict, weight_bnc=l1_cost_volume_w, mask_valid=l1_mask_warped) # B 1 C
l1_points_f1_new_big = self.conv3_l1(l1_cost_volume_sum)
l1_points_f1_new_q = F.dropout(l1_points_f1_new_big, p=0.5, training=self.training)
l1_points_f1_new_t = F.dropout(l1_points_f1_new_big, p=0.5, training=self.training)
l1_q_det = self.conv1_l1(l1_points_f1_new_q)
l1_q_det = l1_q_det / (torch.sqrt(torch.sum(l1_q_det * l1_q_det, dim=-1, keepdim=True) + 1e-10) + 1e-10)
l1_q_det_inv = inv_q(l1_q_det, batch_size)
l1_t_det = self.conv2_l1(l1_points_f1_new_t)
l1_t_coarse_trans = torch.cat([torch.zeros([batch_size, 1, 1]).cuda(), l1_t_coarse], dim=-1)
l1_t_coarse_trans = mul_q_point(l1_q_det, l1_t_coarse_trans, batch_size)
l1_t_coarse_trans = torch.index_select(mul_point_q(l1_t_coarse_trans, l1_q_det_inv, batch_size), 2,
torch.LongTensor(range(1, 4)).cuda())
l1_q = torch.squeeze(mul_point_q(l1_q_det, l1_q_coarse, batch_size), dim=1)
l1_t = torch.squeeze(l1_t_coarse_trans + l1_t_det, dim=1)
# print('l1_refine_time--------', time.time() - start_l1_refine)
################# layer0
# start_l0_refine = time.time()
l0_q_coarse = torch.reshape(l1_q, [batch_size, 1, -1])
l0_t_coarse = torch.reshape(l1_t, [batch_size, 1, -1])
l0_q_inv = inv_q(l0_q_coarse, batch_size)
############# warp layer2 pose
l0_xyz_f1 = torch.reshape(l0_xyz_proj_f1, [batch_size, -1, 3])
l0_xyz_bnc_q = torch.cat([torch.zeros([batch_size, self.out_H_list[0] * self.out_W_list[0], 1]).cuda(), l0_xyz_f1], dim=-1)
l0_flow_warped = mul_q_point(l0_q_coarse, l0_xyz_bnc_q, batch_size)
l0_flow_warped = torch.index_select(mul_point_q(l0_flow_warped, l0_q_inv, batch_size), 2, torch.LongTensor(range(1, 4)).cuda()) + l0_t_coarse
l0_mask = torch.any(l0_xyz_f1 != 0, dim=-1, keepdim=True).to(torch.float32)
l0_flow_warped = l0_flow_warped * l0_mask
########## re-project
l0_xyz_warp_proj_f1, l0_points_warp_proj_f1 = ProjectPCimg2SphericalRing(l0_flow_warped, l0_points_f1, self.out_H_list[0], self.out_W_list[0]) #
l0_xyz_warp_f1 = torch.reshape(l0_xyz_warp_proj_f1, [batch_size, -1, 3])
l0_points_warp_f1 = torch.reshape(l0_points_warp_proj_f1, [batch_size, self.out_H_list[0] * self.out_W_list[0], -1])
l0_mask_warped = torch.any(l0_xyz_warp_f1 != 0, dim=-1, keepdim=False)
l0_mask_warped_proj = torch.reshape(l0_mask_warped, [batch_size, self.out_H_list[0], self.out_W_list[0], -1])
# get the cost volume of warped layer0 flow and the points of frame2
l0_points_warp_proj_f1 = torch.reshape(l0_points_warp_f1, [batch_size, self.out_H_list[0], self.out_W_list[0], -1])
l0_points_proj_f2 = torch.reshape(l0_points_f2, [batch_size, self.out_H_list[0], self.out_W_list[0], -1])
l0_cost_volume = self.cost_volume4(l0_xyz_warp_proj_f1, l0_xyz_proj_f2, l0_points_warp_proj_f1, l0_points_proj_f2)
l0_cost_volume_w_upsample = self.set_upconv3_w_upsample(l0_xyz_warp_proj_f1, l1_xyz_warp_proj_f1, l0_points_warp_proj_f1, l1_cost_volume_w_proj)
l0_cost_volume_upsample = self.set_upconv3_upsample(l0_xyz_warp_proj_f1, l1_xyz_warp_proj_f1, l0_points_warp_proj_f1, l1_cost_volume_proj)
l0_cost_volume_predict = self.flow_predictor3_predict(l0_points_warp_f1, l0_cost_volume_upsample, l0_cost_volume)
l0_cost_volume_w = self.flow_predictor3_w(l0_points_warp_f1, l0_cost_volume_w_upsample, l0_cost_volume_predict)
l0_cost_volume_sum = softmax_valid(feature_bnc=l0_cost_volume_predict, weight_bnc=l0_cost_volume_w,
mask_valid=l0_mask_warped) # B 1 C
l0_points_f1_new_big = self.conv3_l0(l0_cost_volume_sum)
l0_points_f1_new_q = F.dropout(l0_points_f1_new_big, p=0.5, training=self.training)
l0_points_f1_new_t = F.dropout(l0_points_f1_new_big, p=0.5, training=self.training)
l0_q_det = self.conv1_l0(l0_points_f1_new_q)
l0_q_det = l0_q_det / (torch.sqrt(torch.sum(l0_q_det * l0_q_det, dim=-1, keepdim=True) + 1e-10) + 1e-10)
l0_q_det_inv = inv_q(l0_q_det, batch_size)
l0_t_det = self.conv2_l0(l0_points_f1_new_t)
l0_t_coarse_trans = torch.cat([torch.zeros([batch_size, 1, 1]).cuda(), l0_t_coarse], dim=-1)
l0_t_coarse_trans = mul_q_point(l0_q_det, l0_t_coarse_trans, batch_size)
l0_t_coarse_trans = torch.index_select(mul_point_q(l0_t_coarse_trans, l0_q_det_inv, batch_size), 2,
torch.LongTensor(range(1, 4)).cuda())
l0_q = torch.squeeze(mul_point_q(l0_q_det, l0_q_coarse, batch_size), dim=1)
l0_t = torch.squeeze(l0_t_coarse_trans + l0_t_det, dim=1)
l0_q_norm = l0_q / (torch.sqrt(torch.sum(l0_q * l0_q, dim=-1, keepdim=True) + 1e-10) + 1e-10)
l1_q_norm = l1_q / (torch.sqrt(torch.sum(l1_q * l1_q, dim=-1, keepdim=True) + 1e-10) + 1e-10)
l2_q_norm = l2_q / (torch.sqrt(torch.sum(l2_q * l2_q, dim=-1, keepdim=True) + 1e-10) + 1e-10)
l3_q_norm = l3_q / (torch.sqrt(torch.sum(l3_q * l3_q, dim=-1, keepdim=True) + 1e-10) + 1e-10)
return l0_q_norm, l0_t, l1_q_norm, l1_t, l2_q_norm, l2_t, l3_q_norm, l3_t, l1_xyz_f1, q_gt, t_gt, self.w_x, self.w_q
def get_loss(l0_q, l0_t, l1_q, l1_t, l2_q, l2_t, l3_q, l3_t, qq_gt, t_gt, w_x, w_q):
t_gt = torch.squeeze(t_gt)
l0_q_norm = l0_q / (torch.sqrt(torch.sum(l0_q * l0_q, dim=-1, keepdim=True) + 1e-10) + 1e-10)
l0_loss_q = torch.mean(torch.sqrt(torch.sum((qq_gt - l0_q_norm) * (qq_gt - l0_q_norm), dim=-1, keepdim=True) + 1e-10))
l0_loss_x = torch.mean(torch.sqrt((l0_t - t_gt) * (l0_t - t_gt) + 1e-10))
l0_loss = l0_loss_x * torch.exp(-w_x) + w_x + l0_loss_q * torch.exp(-w_q) + w_q
l1_q_norm = l1_q / (torch.sqrt(torch.sum(l1_q * l1_q, -1, keepdim=True) + 1e-10) + 1e-10)
l1_loss_q = torch.mean(torch.sqrt(torch.sum((qq_gt - l1_q_norm) * (qq_gt - l1_q_norm), -1, keepdim=True) + 1e-10))
l1_loss_x = torch.mean(torch.sqrt((l1_t - t_gt) * (l1_t - t_gt) + 1e-10))
l1_loss = l1_loss_x * torch.exp(-w_x) + w_x + l1_loss_q * torch.exp(-w_q) + w_q
l2_q_norm = l2_q / (torch.sqrt(torch.sum(l2_q * l2_q, -1, keepdim=True) + 1e-10) + 1e-10)
l2_loss_q = torch.mean(torch.sqrt(torch.sum((qq_gt - l2_q_norm) * (qq_gt - l2_q_norm), -1, keepdim=True) + 1e-10))
l2_loss_x = torch.mean(torch.sqrt((l2_t - t_gt) * (l2_t - t_gt) + 1e-10))
l2_loss = l2_loss_x * torch.exp(-w_x) + w_x + l2_loss_q * torch.exp(-w_q) + w_q
l3_q_norm = l3_q / (torch.sqrt(torch.sum(l3_q * l3_q, -1, keepdim=True) + 1e-10) + 1e-10)
l3_loss_q = torch.mean(torch.sqrt(torch.sum((qq_gt - l3_q_norm) * (qq_gt - l3_q_norm), -1, keepdim=True) + 1e-10))
l3_loss_x = torch.mean(torch.sqrt((l3_t - t_gt) * (l3_t - t_gt) + 1e-10))
l3_loss = l3_loss_x * torch.exp(-w_x) + w_x + l3_loss_q * torch.exp(-w_q) + w_q
loss_sum = 1.6 * l3_loss + 0.8 * l2_loss + 0.4 * l1_loss + 0.2 * l0_loss
return loss_sum
# if __name__ == "__main__":