-
Notifications
You must be signed in to change notification settings - Fork 5
/
SscDataLoader.py
1260 lines (1146 loc) · 61.6 KB
/
SscDataLoader.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
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
#!/usr/bin/env python
# -*- coding: UTF-8 -*-
"""
SSC DataLoader
---
Jie Li
jieli_cn@163.com
Nanjing University of Science and Technology
University of Adelaide
18/11/2018
"""
import glob
import numpy as np
import numpy.matlib
import torch.utils.data
# from scipy import misc
import imageio
from torchvision import transforms
# import datetime
"""
1.注意 Channel first 和 Channel last
2.注意 D H W 和 W H D 的顺序，
D H W是按照网络中3D_conv的顺序, 本文件中，除了网络的输入输出部分是D H W，其余均按照 W H D。
"""
# data_type 统一为 float32 int32
# TODO: put these in the class
H, W = 480, 640 # frame_height, frame_width, h x w = 480x640
voxel_SIZE = (240, 144, 240) # 240x144x240 = 8294400
voxel_UNIT = 0.02 # 0.02m, length of each grid == 20mm
# vox_margin = 0.24
depth_T_min = 0.5
depth_T_max = 4.8
cam_K = [[518.8579, 0, 320], # K is [fx 0 cx; 0 fy cy; 0 0 1];
[0, 518.8579, 240], # cx = K(1,3); cy = K(2,3);
[0, 0, 1]] # fx = K(1,1); fy = K(2,2);
# C_NUM = 12 # number of classes, 0 - 11, 12 classes
# 'empty','ceiling','floor','wall','window','chair','bed','sofa','table','tvs','furn','objs'
# 12, 'Accessible area'
# 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
seg_class_map = [0, 1, 2, 3, 4, 11, 5, 6, 7, 8, 8, 10, 10, 10, 11, 11, 9, 8, 11, 11, 11,
11, 11, 11, 11, 11, 11, 10, 10, 11, 8, 10, 11, 9, 11, 11, 11] # 0 - 11
# 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
# ssc: color map
colorMap = np.array([[22, 191, 206], # 0 empty, free space
[214, 38, 40], # 1 ceiling
[43, 160, 4], # 2 floor
[158, 216, 229], # 3 wall
[114, 158, 206], # 4 window
[204, 204, 91], # 5 chair new: 180, 220, 90
[255, 186, 119], # 6 bed
[147, 102, 188], # 7 sofa
[30, 119, 181], # 8 table
[188, 188, 33], # 9 tvs
[255, 127, 12], # 10 furn
[196, 175, 214], # 11 objects
[153, 153, 153], # 12 label==255, ignore
]).astype(np.int32)
class NYUv2Dataset(torch.utils.data.Dataset):
def __init__(self, root, mode, encoding='TSDF', downsample=1, data_augment=True):
r"""
Arguments:
root (str): path of folder containing sample data.
mode (str): 'TRAIN', 'VAL', 'TEST', 'TEST', 'VIS'
encoding (str): 'BINARY', 0 for empty, 1 for occupancy
'STSDF', 1 surface, 0 front and empty, -1 behind
'SDF', -1 <= TSDF <= 1, 1 for empty
'NUMBER', NUMBER of points in the bin, 0 for empty
downsample (int): downsample the targets
"""
mode_list = ['TRAIN', 'TRAIN_GN', 'TEST', 'PREDICT', 'TEST_TSDF', 'PREDICT_TSDF']
encoding_list = ['RGB', 'XYZRGB', 'BINARY', 'XYZ', 'STSDF', 'TSDF']
if root is None:
raise Exception("Oops! 'root' is None, please set the right file path.")
self.filepaths = list()
if isinstance(root, list): # 将多个root
for root_i in root:
fp = glob.glob(root_i + '/*.bin')
fp.sort()
self.filepaths.extend(fp)
elif isinstance(root, str):
# print('root is string ', root)
self.filepaths = glob.glob(root + '/*.bin') # List all files in data folder
self.filepaths.sort()
# np.savetxt('NYUCADtrain_filenames.txt', self.filepaths, fmt="%s")
self.mode = mode
self.encoding = encoding
self.data_augment = False
print('self.data_augment', self.data_augment)
# Converts a PIL Image or numpy.ndarray (H x W x C) in the range [0, 255] \
# to a torch.FloatTensor of shape (C x H x W) in the range [0.0, 1.0].
self.transforms_rgb = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
if mode not in mode_list:
raise Exception("Oops! '{}' is not a valid mode. Try {}.".format(mode, mode_list))
if encoding not in encoding_list:
raise Exception("Oops! '{}' is not a valid Encoding. Try {}.".format(encoding, encoding_list))
if len(self.filepaths) == 0:
raise Exception("Oops! That was no valid data in '{}'.".format(root))
if not isinstance(downsample, int):
raise Exception("An int is expected, but got {}".format(type(downsample)))
self.downsample = downsample # int, downsample = 4, in labeled data, get 1 voxel from each 4
self.vox_size = (voxel_SIZE[0] / downsample, voxel_SIZE[1] / downsample, voxel_SIZE[2] / downsample)
print('SuncgDataset:mode:{}, encoding:{}, {} files, Resolution:{}'.format(mode, encoding, len(self.filepaths), self.vox_size))
def __getitem__(self, index):
r"""
Shape:
voxels: (1, 240, 144, 240), pytorch needs channel-first
target: (1, 60, 36, 60)
"""
_name = self.filepaths[index][:-4]
depth = self._read_depth(_name + '.png', 480, 640) # (h, w)
vox_origin, cam_pose, rle = self._read_rle(_name + '.bin')
target_hr = self._rle2voxel(rle, _name + '.bin') if self.mode is not 'PREDICT' else None
# target_lr = self._downsample_label(target_hr, 4) if self.mode is not 'PREDICT' else None
# TODO rgb vs depth
# binary_vox, xyz_vox = self._depth2voxel(depth, cam_pose, vox_origin, unit=voxel_UNIT)
# voxels = rgb_vox if self.encoding == 'RGB' else xyzrgb_vox
# TODO xyzrgb/240 144 240 255 255 255
if self.encoding == 'BINARY':
binary_vox, _, _ = self._depth2voxel(depth, cam_pose, vox_origin, unit=voxel_UNIT)
voxels = binary_vox.reshape(binary_vox.shape + (1,)) # (W, H, D, 1)
if self.encoding == 'XYZ':
binary_vox, xyz_vox, _ = self._depth2voxel(depth, cam_pose, vox_origin, unit=voxel_UNIT)
# voxels = xyz_vox # (W, H, D, ３), 0-240, 0-144, 0-240
voxels = xyz_vox / voxel_SIZE # (W, H, D, ３), 0-1. # TODO type np.float32
elif self.encoding == 'STSDF':
binary_vox, _, _ = self._depth2voxel(depth, cam_pose, vox_origin, unit=voxel_UNIT)
stsdf_hr = self._get_stsdf(depth, binary_vox, vox_origin, cam_pose) # (W, H, D)
voxels = stsdf_hr.reshape(stsdf_hr.shape + (1,)) # (W, H, D, 1)
elif self.encoding == 'TSDF':
binary_vox, _, position = self._depth2voxel(depth, cam_pose, vox_origin, unit=voxel_UNIT)
npz_file = np.load(_name + '.npz')
# target_lr = npz_file['label'].astype(np.int32) # (W, H, D), SUNCG
tsdf_hr = npz_file['tsdf'] # SUNCG (W, H, D)
voxels = tsdf_hr.reshape(tsdf_hr.shape + (1,)) # (W, H, D, 1)
elif self.encoding == 'RGB':
rgb = self._read_rgb(_name + '.jpg', 480, 640) # (h=image_height, w=image_width, 3)
rgb = rgb / 255.0 # if normalise else rgb # normalise rgb from [0, 255] to [0, 1]
binary_vox, rgb_vox, xyz_vox = self._rgbd2voxel(rgb, depth, cam_pose, vox_origin, voxel_UNIT)
voxels = rgb_vox
elif self.encoding == 'XYZRGB':
rgb = self._read_rgb(_name + '.jpg', 480, 640) # (h=image_height, w=image_width, 3)
rgb = rgb / 255.0 # if normalise else rgb # normalise rgb from [0, 255] to [0, 1]
binary_vox, rgb_vox, xyz_vox = self._rgbd2voxel(rgb, depth, cam_pose, vox_origin, voxel_UNIT)
xyz_vox = xyz_vox / voxel_SIZE # xyz_vox.shape[:3]
voxels = np.concatenate((xyz_vox, rgb_vox), 3) # (W, H, D, 6)
if self.mode == 'TRAIN': # Have labeled data
# ---- data augment
if self.data_augment:
voxels, target_hr = self._data_augment(voxels, target_hr, self.encoding)
target_lr = self._downsample_label(target_hr, 4)
# Save GT
# ply_filename = _name + '-HR_GT.ply'
# self.labeled_voxel2ply(target_hr, ply_filename)
# Save depth_voxel
# ply_filename = _name + '-LR_depth_voxel_method2-2.ply'
# self._depth_voxel2ply(binary_vox, ply_filename, encoding='BINARY')
# if self.downsample == 1: # HR
# one_hot_target_vox = one_hot_embedding(target_hr, 12) # W H D C. for GAN
# return voxels.T, target_hr.T, target_lr.T, one_hot_target_vox.T, _name + '.png' # C D H W, for Conv3d
#
# if self.downsample == 4: # LR
# one_hot_target_vox = one_hot_embedding(target_lr, 12) # W H D C. for GAN
# return voxels.T, target_lr.T, one_hot_target_vox.T, _name + '.png' # C D H W, for Conv3d
depth = depth.reshape((1,) + depth.shape)
# rgb = self._read_rgb(_name + '.jpg', 480, 640) # (h=image_height, w=image_width, 3)
# rgb_tesnor = self.transforms_rgb(rgb) # (C x H x W) = (3, 480, 640)
if self.downsample == 1: # HR
# one_hot_target_vox = one_hot_embedding(target_hr, 12) # W H D C. for GAN
# gn = self._get_gradient_norm_semantic(target_hr, d=2) # (W H D)
# gn = gn + 0.2
# 注意position[H, W, 3] 对应X，Y，Z，而voxels的坐标顺序为Z,Y,X
return depth, voxels.T, target_hr.T, target_lr.T, position, _name + '.png' # C D H W, for Conv3d
if self.downsample == 4: # LR
# ply_filename = _name + '-HR_RGBvox-all.ply'
# self.rgb_voxel2ply(rgb_vox, ply_filename)
# one_hot_target_vox = one_hot_embedding(target_lr, 12) # W H D C. for GAN
# gn = self._get_gradient_norm_semantic(target_lr, d=2) # (W H D)
# print('np.amin(gn), np.amax(gn)', np.amin(gn), np.amax(gn))
# self._gn_voxel2ply(gn, _name + 'gn.ply')
# gn = gn + 1 # TODO 0.2 比较的参数
# gn[target_lr == 255] = 0
# return voxels.T, target_lr.T, gn.T, _name + '.png' # C D H W, for Conv3d
# 注意position[H, W, 3] 对应X，Y，Z，而voxels的坐标顺序为Z,Y,X
return depth, voxels.T, target_lr.T, position, _name + '.png' # C D H W, for Conv3d
# return depth, rgb_tesnor, voxels.T, target_lr.T, position, _name + '.png' # C D H W, for Conv3d
# return rgb_tesnor, voxels.T, target_lr.T, position, _name + '.png' # C D H W, for Conv3d
if self.mode == 'TRAIN_GN': # Have labeled data
# ---- data augment
if self.data_augment:
voxels, target_hr = self._data_augment(voxels, target_hr, self.encoding)
target_lr = self._downsample_label(target_hr, 4)
if self.downsample == 1: # HR
gn_hr = self._get_gradient_norm_semantic(target_hr, d=2, ord=np.inf) # (W H D) TODO ord=1, 2, inf
gn_hr = gn_hr + 1 #
gn_lr = self._get_gradient_norm_semantic(target_lr, d=1, ord=np.inf) # (W H D)
gn_lr = gn_lr + 1 # 1 为原始权重，各类别、各体素权重均为1
return voxels.T, target_hr.T, target_lr.T, gn_hr.T, gn_lr.T, _name + '.png' # C D H W, for Conv3d
if self.downsample == 4: # LR
# ply_filename = _name + '-HR_RGBvox-all.ply'
# self.rgb_voxel2ply(rgb_vox, ply_filename)
# one_hot_target_vox = one_hot_embedding(target_lr, 12) # W H D C. for GAN
gn_lr = self._get_gradient_norm_semantic(target_lr, d=2, ord=2) # (W H D)
gn_lr = gn_lr + 1 # 1 为原始权重，各类别、各体素权重均为1
return voxels.T, target_lr.T, gn_lr.T, _name + '.png' # C D H W, for Conv3d
if self.mode == 'TEST': # Have labeled data, nonempty=None
target = self._downsample_label(target_hr, self.downsample)
return voxels.T, target.T, _name + '.png' # C D H W, for Conv3d
if self.mode == 'TEST_TSDF': # Have labeled data
target = self._downsample_label(target_hr, self.downsample)
# ---- STSDF
# stsdf_hr = self._get_stsdf(depth, binary_vox, vox_origin, cam_pose)
# stsdf = self._downsample_stsdf(stsdf_hr, self.downsample)
# nonempty = self.get_nonempty(stsdf, 'STSDF') # 更合理， 差别在于天花板顶上与墙后的处理
# # nonempty = self.get_nonempty2(stsdf, target, 'STSDF') # 这个更符合SUNCG的做法
# ---- TSDF
if self.encoding != 'TSDF':
npz_file = np.load(_name + '.npz')
tsdf_hr = npz_file['tsdf'] # SUNCG (W, H, D)
tsdf = self._downsample_tsdf(tsdf_hr, self.downsample)
# nonempty = self.get_nonempty(tsdf, 'TSDF')
nonempty = self.get_nonempty2(tsdf, target, 'TSDF') # 这个更符合SUNCG的做法
# ---- save ply
# ply_filename = _name + '-HR_empty-SUNCG-STSDF.ply'
# self._depth_voxel2ply(nonempty, ply_filename, encoding='EMPTY')
# ply_filename = _name + '-HR_binary-SUNCG.ply'
# binary_vox, xyz_vox = self._depth2voxel(depth, cam_pose, vox_origin, unit=voxel_UNIT)
# self._depth_voxel2ply(binary_vox, ply_filename, encoding='BINARY')
# ply_filename = _name + '-HR_stsdf-SUNCG-surface.ply'
# SuncgDataset._depth_voxel2ply(stsdf_hr, ply_filename, encoding='STSDF')
depth = depth.reshape((1,) + depth.shape)
# rgb = self._read_rgb(_name + '.jpg', 480, 640) # (h=image_height, w=image_width, 3)
# rgb_tesnor = self.transforms_rgb(rgb) # (C x H x W) = (3, 480, 640)
# return depth, rgb_tesnor, voxels.T, target.T, nonempty.T, position, _name + '.png' # C D H W, for Conv3d
# return rgb_tesnor, voxels.T, target.T, nonempty.T, position, _name + '.png'
return depth, voxels.T, target.T, nonempty.T, position, _name + '.png'
if self.mode == 'PREDICT': # Do not have labeled data, only show predict results
return voxels.T, _name + '.png' # C D H W, for Conv3d
if self.mode == 'PREDICT_TSDF': # Have labeled data
stsdf_hr = self._get_stsdf(depth, binary_vox, vox_origin, cam_pose)
stsdf = self._downsample_stsdf(stsdf_hr, self.downsample)
nonempty = self.get_nonempty(stsdf, 'STSDF')
# nonempty = self.get_nonempty2(stsdf, target, 'STSDF')
return voxels.T, nonempty.T, _name + '.png' # C D H W, for Conv3d
def __len__(self):
return len(self.filepaths)
@classmethod
def _data_augment(cls, voxels, target, encoding):
# encoding == 'RGB', 'XYZRGB', 'BINARY', 'XYZ', 'STSDF': _empty = 0
# encoding == 'TSDF': _empty = np.float32(0.001)
_empty = np.float32(0.001) if encoding == 'TSDF' else 0
if np.random.rand() < 0.8: # 沿Z轴(对应D),往后(Z变大的方向)平移, 最多平移Z轴的20%，平移后，空出来的部分置为空
sz = int(np.random.rand() * voxels.shape[2] * 0.2) + 1 # at least, move 1 grid
voxels = np.roll(voxels, shift=sz, axis=2) # Roll array elements along axis Z.
target = np.roll(target, shift=sz, axis=2)
voxels[:, :, :sz, :] = _empty
target[:, :, :sz] = 255 # 0, empty --> 255, ignore
if np.random.rand() < 0.8: # 沿x轴(对应W),往x变大的方向平移, 最多平移X轴的10%，平移后，空出来的部分置为空
move_opposite = True if np.random.rand() < 0.5 else False # move direction
sx_1 = int(np.random.rand() * voxels.shape[0] * 0.1) + 1 # at least, move 1 grid
sx_1 = voxels.shape[0] - sx_1 if move_opposite else sx_1
voxels = np.roll(voxels, shift=sx_1, axis=0) # Roll array elements along axis X.
target = np.roll(target, shift=sx_1, axis=0)
if move_opposite: # 往x变小的方向平移
voxels[sx_1:, :, :, :] = _empty
target[sx_1:, :, :] = 255
else: # 往x变大的方向平移
voxels[:sx_1, :, :, :] = _empty
target[:sx_1, :, :] = 255
if np.random.rand() < 0.5: # 左右反转
target = np.flip(target, 0).copy() # Flip an array vertically (axis=0).
voxels = np.flip(voxels, 0).copy()
return voxels, target
@staticmethod
def _get_gradient_norm(data, d=1, ord=3): # TODO d=1, ord=1 作为实验比较的参数
""" 依据GT计算每个点的梯度，用梯度的大小作为loss的权重""" #
# data [W, H, D], Ground truth, array_like. An N-dimensional array containing samples of a scalar function.
# d: single scalar to specify a sample distance for all dimensions.
f = np.zeros(data.shape, dtype=np.float32)
f[data > 0] = 1
g = np.gradient(f, d) # list of ndarray, (3, [W, H, D])
g = np.asarray(g) # [3, W, H, D]
gn = np.linalg.norm(g, ord=ord, axis=0) # [W, H, D]
return gn
@staticmethod
def _get_gradient_norm_semantic(data, d=1, ord=2): # TODO d=1, ord=1 作为实验比较的参数
""" 依据GT计算每个点的梯度，用梯度的大小作为loss的权重,
d: single scalar to specify a sample distance for all dimensions.
ord: Order of the norm
""" #
# data [W, H, D]
c = 12
f = np.zeros(data.shape, dtype=np.float32)
gn = np.zeros(data.shape, dtype=np.float32)
# data_tmp = data
data_tmp = np.copy(data)
data_tmp[data == 255] = 0 # ingnore_index=255, treat these voxels as free space
for idx in range(c):
f.fill(0)
f[data_tmp == idx] = 1
g = np.gradient(f, d) # list of ndarray, (3, [W, H, D])
g = np.asarray(g) # [3, W, H, D]
gn += np.linalg.norm(g, ord=ord, axis=0) # [W, H, D]
# print(type(gn), type(gn[0,0,0]))
return gn
@staticmethod
def _read_depth(depth_filename, img_h=480, img_w=640):
r"""Read a depth image with size H x W
and save the depth values (in millimeters) into a 2d numpy array.
The depth image file is assumed to be in 16-bit PNG format, depth in millimeters.
"""
# depth = misc.imread(depth_filename) / 8000.0 # numpy.float64
# depth = misc.imresize(depth, (img_h, img_w)) # numpy.uint8
depth = imageio.imread(depth_filename) / 8000.0 # numpy.float64
depth = np.asarray(depth)
assert depth.shape == (img_h, img_w), 'incorrect default size'
return depth
@staticmethod
def _read_rgb(rgb_filename, img_h=480, img_w=640): # 0.01s
r"""Read a RGB image with size H x W
"""
# rgb = misc.imread(rgb_filename) # <type 'numpy.ndarray'>, numpy.uint8, (480, 640, 3)
# rgb = misc.imresize(rgb, (img_h, img_w)) # (H, W, 3)
rgb = imageio.imread(rgb_filename) # <type 'numpy.ndarray'>, numpy.uint8, (480, 640, 3)
rgb = np.asarray(rgb)
# rgb = np.rollaxis(rgb, 2, 0) # (H, W, 3)-->(3, H, W)
return rgb
@staticmethod
def _read_rle(rle_filename): # 0.0005s
r"""Read RLE compression data
Return:
vox_origin,
cam_pose,
vox_rle, voxel label data from file
Shape:
vox_rle, (240, 144, 240), if downsample==4, then (60, 36, 60)
"""
fid = open(rle_filename, 'rb')
vox_origin = np.fromfile(fid, np.float32, 3).T # Read voxel origin in world coordinates
cam_pose = np.fromfile(fid, np.float32, 16).reshape((4, 4)) # Read camera pose
vox_rle = np.fromfile(fid, np.uint32).reshape((-1, 1)).T # Read voxel label data from file
vox_rle = np.squeeze(vox_rle) # 2d array: (1 x N), to 1d array: (N , )
fid.close()
return vox_origin, cam_pose, vox_rle
@staticmethod
def _get_xyz(size=voxel_SIZE):
"""x 水平 y高低 z深度"""
_x = np.zeros(size, dtype=np.int32)
_y = np.zeros(size, dtype=np.int32)
_z = np.zeros(size, dtype=np.int32)
for i_h in range(size[0]): # x, y, z
_x[i_h, :, :] = i_h # x, left-right flip
for i_w in range(size[1]):
_y[:, i_w, :] = i_w # y, up-down flip
for i_d in range(size[2]):
_z[:, :, i_d] = i_d # z, front-back flip
return _x, _y, _z
@classmethod
def _get_stsdf(cls, depth, voxel_binary, vox_origin, cam_pose):
r"""simplified TSDF, encoding == 'STSDF'
Shape:
depth: (H, W)
voxel_binary: (240, 144, 240)
voxel_stsdf: (240, 144, 240)
"""
c = cam_pose
p_base = np.zeros(voxel_SIZE + (3,), dtype=np.float32) # points
p_cam2 = np.zeros(voxel_SIZE + (3,), dtype=np.float32) # points in camera view
pixel_xy = np.zeros(voxel_SIZE + (2,), dtype=np.int32)
_x, _y, _z = cls._get_xyz()
p_base[:, :, :, 0] = _x * voxel_UNIT + vox_origin[0]
p_base[:, :, :, 1] = _z * voxel_UNIT + vox_origin[1]
p_base[:, :, :, 2] = _y * voxel_UNIT + vox_origin[2]
p_base[:, :, :, 0] = p_base[:, :, :, 0] - c[0][3]
p_base[:, :, :, 1] = p_base[:, :, :, 1] - c[1][3]
p_base[:, :, :, 2] = p_base[:, :, :, 2] - c[2][3]
p_cam2[:, :, :, 0] = c[0][0] * p_base[:, :, :, 0] + c[1][0] * p_base[:, :, :, 1] + c[2][0] * p_base[:, :, :, 2]
p_cam2[:, :, :, 1] = c[0][1] * p_base[:, :, :, 0] + c[1][1] * p_base[:, :, :, 1] + c[2][1] * p_base[:, :, :, 2]
p_cam2[:, :, :, 2] = c[0][2] * p_base[:, :, :, 0] + c[1][2] * p_base[:, :, :, 1] + c[2][2] * p_base[:, :, :, 2]
# NYUCADtrain NYU0601_0000.png, p_cam2[19 99 7 2]==0
# NYUCADtest NYU0761_0000.png
if np.count_nonzero(p_cam2[:, :, :, 2] == 0):
tt_idx = np.nonzero(p_cam2[:, :, :, 2] == 0)
tt_idx = np.stack(tt_idx, axis=1)
b = np.ones((tt_idx.shape[0], 1), dtype=np.int32) * 2
tt_idx = np.concatenate((tt_idx, b), axis=1)
p_cam2[tt_idx] += 0.0001
# print(f, '/0 bug')
pixel_xy[:, :, :, 0] = np.round(cam_K[0][0] * (p_cam2[:, :, :, 0] / p_cam2[:, :, :, 2]) + cam_K[0][2])
pixel_xy[:, :, :, 1] = np.round(cam_K[1][1] * (p_cam2[:, :, :, 1] / p_cam2[:, :, :, 2]) + cam_K[1][2])
# initial to empty
# 1 ----- 0 empty
STSDF_EMPTY, STSDF_SURFACE, STSDF_OCCLUD = 0, 1, -1
voxel_stsdf = np.zeros(voxel_SIZE, dtype=np.float32)
# 2 ----- 1 empty
# STSDF_EMPTY, STSDF_SURFACE, STSDF_OCCLUD = 1, 0, -1
# voxel_stsdf = np.ones(voxel_SIZE, dtype=np.float32)
idx_h = np.rint(pixel_xy[:, :, :, 1]).astype(np.int32)
idx_w = np.rint(pixel_xy[:, :, :, 0]).astype(np.int32)
idx_h[idx_h >= 480] = 0 # 超出场景范围
idx_w[idx_w >= 640] = 0
idx_h[idx_h < 0] = 0
idx_w[idx_w < 0] = 0
pt_depth = depth[idx_h, idx_w]
voxel_stsdf[pt_depth[:] <= p_cam2[:, :, :, 2]] = STSDF_OCCLUD # 被遮挡
# voxel_stsdf[abs(pt_depth - p_cam2[:, :, :, 2]) < 0.0001] = 1 # 超出搜索范围
voxel_stsdf[np.round(pt_depth) == 0] = STSDF_EMPTY
# if pt_depth < depth_T_min or pt_depth > depth_T_max: # 超出场景范围
# voxel_stsdf[x, y, z] = 1
idx_h = pixel_xy[:, :, :, 1].astype(np.int32)
idx_w = pixel_xy[:, :, :, 0].astype(np.int32)
voxel_stsdf[idx_h >= 480] = STSDF_EMPTY # 超出场景范围
voxel_stsdf[idx_w >= 640] = STSDF_EMPTY
voxel_stsdf[idx_h < 0] = STSDF_EMPTY
voxel_stsdf[idx_w < 0] = STSDF_EMPTY
voxel_stsdf[p_cam2[:, :, :, 2] <= 0] = STSDF_EMPTY # 视野之外
voxel_stsdf[voxel_binary > 0] = STSDF_SURFACE # surface
# 1 ----- 0 empty
# voxel_stsdf = 1 - voxel_stsdf # surface, empty, occulted: 1, 0, -1 ---> 0 , 1 , -2
# 2 ----- 1 empty
# voxel_stsdf = 1 - voxel_stsdf # surface, empty, occulted: 0, 1, -1 ---> 1 , 0 , -2
del p_base, p_cam2, pixel_xy
return voxel_stsdf
@classmethod
def _rle2voxel(cls, rle, rle_filename=''):
r"""Read voxel label data from file (RLE compression), and convert it to fully occupancy labeled voxels.
In the data loader of pytorch, only single thread is allowed.
For multi-threads version and more details, see 'readRLE.py'.
output: seg_label: 3D numpy array, size 240 x 144 x 240
"""
# ---- Read RLE
# vox_origin, cam_pose, rle = cls._read_rle(rle_filename)
# ---- Uncompress RLE, 0.9s
seg_label = np.zeros(voxel_SIZE[0] * voxel_SIZE[1] * voxel_SIZE[2], dtype=np.uint8) # segmentation label
vox_idx = 0
for idx in range(int(rle.shape[0] / 2.0)):
check_val = rle[idx * 2]
check_iter = rle[idx * 2 + 1]
if check_val >= 37 and check_val != 255: # 37 classes to 12 classes
print('RLE {} check_val: {}'.format(rle_filename, check_val))
# seg_label_val = 1 if check_val < 37 else 0 # 37 classes to 2 classes: empty or occupancy
# seg_label_val = 255 if check_val == 255 else seg_class_map[check_val]
seg_label_val = seg_class_map[check_val] if check_val != 255 else 255 # 37 classes to 12 classes
seg_label[vox_idx: vox_idx+check_iter] = np.matlib.repmat(seg_label_val, 1, check_iter)
vox_idx = vox_idx + check_iter
seg_label = seg_label.reshape(voxel_SIZE) # 3D array, size 240 x 144 x 240
return seg_label
@classmethod # method 1
def _depth2voxel_old(cls, depth, cam_pose, vox_origin, unit=0.02):
# ---- Get point in camera coordinate
gx, gy = np.meshgrid(range(W), range(H))
pt_cam = np.zeros((H, W, 3), dtype=np.float32)
pt_cam[:, :, 0] = (gx - cam_K[0][2]) * depth / cam_K[0][0] # x
pt_cam[:, :, 1] = (gy - cam_K[1][2]) * depth / cam_K[1][1] # y
pt_cam[:, :, 2] = depth # z, in meter
# ---- Get point in world coordinate
p = cam_pose
pt_world = np.zeros((H, W, 3), dtype=np.float32)
pt_world[:, :, 0] = p[0][0] * pt_cam[:, :, 0] + p[0][1] * pt_cam[:, :, 1] + p[0][2] * pt_cam[:, :, 2] + p[0][3]
pt_world[:, :, 1] = p[1][0] * pt_cam[:, :, 0] + p[1][1] * pt_cam[:, :, 1] + p[1][2] * pt_cam[:, :, 2] + p[1][3]
pt_world[:, :, 2] = p[2][0] * pt_cam[:, :, 0] + p[2][1] * pt_cam[:, :, 1] + p[2][2] * pt_cam[:, :, 2] + p[2][3]
pt_world[:, :, 0] = pt_world[:, :, 0] - vox_origin[0]
pt_world[:, :, 1] = pt_world[:, :, 1] - vox_origin[1]
pt_world[:, :, 2] = pt_world[:, :, 2] - vox_origin[2]
# ---- Aline the coordinates with labeled data (RLE .bin file)
pt_world2 = np.zeros(pt_world.shape, dtype=np.float32) # (h, w, 3)
# pt_world2 = pt_world
pt_world2[:, :, 0] = pt_world[:, :, 0] # x 水平
pt_world2[:, :, 1] = pt_world[:, :, 2] # y 高低
pt_world2[:, :, 2] = pt_world[:, :, 1] # z 深度
# pt_world2[:, :, 0] = pt_world[:, :, 1] # x 原始paper方法
# pt_world2[:, :, 1] = pt_world[:, :, 2] # y
# pt_world2[:, :, 2] = pt_world[:, :, 0] # z
# ---- World coordinate to grid/voxel coordinate
point_grid = pt_world2 / unit # Get point in grid coordinate, each grid is a voxel
point_grid = np.rint(point_grid).astype(np.int32).reshape((-1, 3)) # (h*w, 3)
# ---- crop depth to grid/voxel
# binary encoding '01': 0 for empty, 1 for occupancy
# voxel_binary = np.zeros(voxel_SIZE, dtype=np.uint8) # (W, H, D)
voxel_binary = np.zeros(voxel_SIZE, dtype=np.float32) # (W, H, D)
voxel_xyz = np.zeros(voxel_SIZE + (3,), dtype=np.float32) # (W, H, D, 3)
for i_idx in range(len(point_grid)):
i_x, i_y, i_z = point_grid[i_idx, :]
# i_x, i_y, i_z = int(i_x), int(i_y), int(i_z)
if i_x < voxel_SIZE[0] and i_y < voxel_SIZE[1] and i_z < voxel_SIZE[2] \
and i_x >= 0 and i_y >= 0 and i_z >= 0:
voxel_binary[i_x][i_y][i_z] = 1 # the bin has at least one point (bin is not empty)
voxel_xyz[i_x, i_y, i_z, :] = point_grid[i_idx, :]
# output --- 3D Tensor, 240 x 144 x 240
del depth, gx, gy, pt_cam, pt_world, pt_world2, point_grid # Release Memory
return voxel_binary, voxel_xyz # (W, H, D), (W, H, D, 3)
@classmethod # method 2, new
def _depth2voxel(cls, depth, cam_pose, vox_origin, unit=0.02):
# ---- Get point in camera coordinate
gx, gy = np.meshgrid(range(W), range(H))
pt_cam = np.zeros((H, W, 3), dtype=np.float32)
pt_cam[:, :, 0] = (gx - cam_K[0][2]) * depth / cam_K[0][0] # x
pt_cam[:, :, 1] = (gy - cam_K[1][2]) * depth / cam_K[1][1] # y
pt_cam[:, :, 2] = depth # z, in meter
# ---- Get point in world coordinate
p = cam_pose
pt_world = np.zeros((H, W, 3), dtype=np.float32)
pt_world[:, :, 0] = p[0][0] * pt_cam[:, :, 0] + p[0][1] * pt_cam[:, :, 1] + p[0][2] * pt_cam[:, :, 2] + p[0][3]
pt_world[:, :, 1] = p[1][0] * pt_cam[:, :, 0] + p[1][1] * pt_cam[:, :, 1] + p[1][2] * pt_cam[:, :, 2] + p[1][3]
pt_world[:, :, 2] = p[2][0] * pt_cam[:, :, 0] + p[2][1] * pt_cam[:, :, 1] + p[2][2] * pt_cam[:, :, 2] + p[2][3]
pt_world[:, :, 0] = pt_world[:, :, 0] - vox_origin[0]
pt_world[:, :, 1] = pt_world[:, :, 1] - vox_origin[1]
pt_world[:, :, 2] = pt_world[:, :, 2] - vox_origin[2]
# ---- Aline the coordinates with labeled data (RLE .bin file)
pt_world2 = np.zeros(pt_world.shape, dtype=np.float32) # (h, w, 3)
# pt_world2 = pt_world
pt_world2[:, :, 0] = pt_world[:, :, 0] # x 水平
pt_world2[:, :, 1] = pt_world[:, :, 2] # y 高低
pt_world2[:, :, 2] = pt_world[:, :, 1] # z 深度
# pt_world2[:, :, 0] = pt_world[:, :, 1] # x 原始paper方法
# pt_world2[:, :, 1] = pt_world[:, :, 2] # y
# pt_world2[:, :, 2] = pt_world[:, :, 0] # z
"""
# ---- World coordinate to grid/voxel coordinate
point_grid = pt_world2 / unit # Get point in grid coordinate, each grid is a voxel
point_grid = np.rint(point_grid).astype(np.int32).reshape((-1, 3)) # (h*w, 3)
# ---- crop depth to grid/voxel
# binary encoding '01': 0 for empty, 1 for occupancy
# voxel_binary = np.zeros(voxel_SIZE, dtype=np.uint8) # (W, H, D)
voxel_binary = np.zeros(voxel_SIZE, dtype=np.float32) # (W, H, D)
voxel_xyz = np.zeros(voxel_SIZE + (3,), dtype=np.float32) # (W, H, D, 3)
position = np.zeros((H, W, 3), dtype=np.int32)
for i_idx in range(len(point_grid)):
i_x, i_y, i_z = point_grid[i_idx, :]
# i_x, i_y, i_z = int(i_x), int(i_y), int(i_z)
if i_x < voxel_SIZE[0] and i_y < voxel_SIZE[1] and i_z < voxel_SIZE[2] \
and i_x >= 0 and i_y >= 0 and i_z >= 0:
voxel_binary[i_x][i_y][i_z] = 1 # the bin has at least one point (bin is not empty)
voxel_xyz[i_x, i_y, i_z, :] = point_grid[i_idx, :]
h = i_idx / W
w = i_idx - h * W
position[h, w, :] = point_grid[i_idx, :] # 记录图片上的每个像素对应的voxel位置
"""
# ---- World coordinate to grid/voxel coordinate
point_grid = pt_world2 / unit # Get point in grid coordinate, each grid is a voxel
point_grid = np.rint(point_grid).astype(np.int32) # .reshape((-1, 3)) # (H*W, 3) (H, W, 3)
# print(point_grid.shape)
# ---- crop depth to grid/voxel
# binary encoding '01': 0 for empty, 1 for occupancy
# voxel_binary = np.zeros(voxel_SIZE, dtype=np.uint8) # (W, H, D)
voxel_binary = np.zeros(voxel_SIZE, dtype=np.float32) # (W, H, D)
voxel_xyz = np.zeros(voxel_SIZE + (3,), dtype=np.float32) # (W, H, D, 3)
# position = np.zeros((H, W, 3), dtype=np.int32)
position = np.zeros((H, W), dtype=np.int32)
# [rows, cols, c] = point_grid.shape # (H, W, 3)
# rows = H
# cols = W
for h in range(H):
for w in range(W):
i_x, i_y, i_z = point_grid[h, w, :]
if 0 <= i_x < voxel_SIZE[0] and 0 <= i_y < voxel_SIZE[1] and 0 <= i_z < voxel_SIZE[2]:
voxel_binary[i_x][i_y][i_z] = 1 # the bin has at least one point (bin is not empty)
voxel_xyz[i_x, i_y, i_z, :] = point_grid[h, w, :]
# position[h, w, :] = point_grid[h, w, :] # 记录图片上的每个像素对应的voxel位置
# idx = np.ravel_multi_index(point_grid[h, w, :], voxel_SIZE)
position[h, w] = np.ravel_multi_index(point_grid[h, w, :], voxel_SIZE) # 记录图片上的每个像素对应的voxel位置
# print(num[h, w])
# output --- 3D Tensor, 240 x 144 x 240
del depth, gx, gy, pt_cam, pt_world, pt_world2, point_grid # Release Memory
# position[h,w]
return voxel_binary, voxel_xyz, position # (W, H, D), (W, H, D, 3)
@classmethod
def _rgbd2voxel(cls, rgb, depth, cam_pose, vox_origin, unit=0.02):
# ---- Get point in camera coordinate
gx, gy = np.meshgrid(range(W), range(H))
pt_cam = np.zeros((H, W, 3), dtype=np.float32)
pt_cam[:, :, 0] = (gx - cam_K[0][2]) * depth / cam_K[0][0] # x
pt_cam[:, :, 1] = (gy - cam_K[1][2]) * depth / cam_K[1][1] # y
pt_cam[:, :, 2] = depth # z, in meter
# ---- Get point in world coordinate
p = cam_pose
pt_world = np.zeros((H, W, 3), dtype=np.float32)
pt_world[:, :, 0] = p[0][0] * pt_cam[:, :, 0] + p[0][1] * pt_cam[:, :, 1] + p[0][2] * pt_cam[:, :, 2] + p[0][3]
pt_world[:, :, 1] = p[1][0] * pt_cam[:, :, 0] + p[1][1] * pt_cam[:, :, 1] + p[1][2] * pt_cam[:, :, 2] + p[1][3]
pt_world[:, :, 2] = p[2][0] * pt_cam[:, :, 0] + p[2][1] * pt_cam[:, :, 1] + p[2][2] * pt_cam[:, :, 2] + p[2][3]
pt_world[:, :, 0] = pt_world[:, :, 0] - vox_origin[0]
pt_world[:, :, 1] = pt_world[:, :, 1] - vox_origin[1]
pt_world[:, :, 2] = pt_world[:, :, 2] - vox_origin[2]
# ---- Aline the coordinates with labeled data (RLE .bin file)
pt_world2 = np.zeros(pt_world.shape, dtype=np.float32) # (h, w, 3)
# pt_world2 = pt_world
pt_world2[:, :, 0] = pt_world[:, :, 0] # x 水平
pt_world2[:, :, 1] = pt_world[:, :, 2] # y 高低
pt_world2[:, :, 2] = pt_world[:, :, 1] # z 深度
# pt_world2[:, :, 0] = pt_world[:, :, 1] # x 原始paper方法
# pt_world2[:, :, 1] = pt_world[:, :, 2] # y
# pt_world2[:, :, 2] = pt_world[:, :, 0] # z
# ---- World coordinate to grid/voxel coordinate
# point_grid = pt_world2 / voxel_UNIT # Get point in grid coordinate, each grid is a voxel
point_grid = pt_world2 / unit # Get point in grid coordinate, each grid is a voxel
point_grid = np.rint(point_grid).astype(int).reshape((-1, 3)) # (h*w, 3)
# point_grid = point_grid.reshape((-1, 3)) # (h*w, 3)
rgb = rgb.reshape((-1, 3)) # (h*w, 3)
# ---- crop depth to grid/voxel
# binary encoding '01': 0 for empty, 1 for occupancy
voxel_binary = np.zeros(voxel_SIZE, dtype=np.float32) # (W, H, D) dtype=np.uint8
voxel_xyz = np.zeros(voxel_SIZE + (3,), dtype=np.float32) # (W, H, D, 3)
voxel_rgb = np.zeros(voxel_SIZE + (3,), dtype=np.float32) # (W, H, D, 3)
# voxel_xyzrgb = np.zeros(voxel_SIZE + (6,), dtype=np.float32) # (W, H, D, 6)
for i_idx in range(len(point_grid)):
i_x, i_y, i_z = point_grid[i_idx, :]
# i_x, i_y, i_z = int(i_x), int(i_y), int(i_z)
if i_x < voxel_SIZE[0] and i_y < voxel_SIZE[1] and i_z < voxel_SIZE[2] \
and i_x >= 0 and i_y >= 0 and i_z >= 0:
# if 0 <= i_x < voxel_SIZE[0] and 0 <= i_y < voxel_SIZE[1] and 0 <= i_z < voxel_SIZE[2]:
voxel_binary[i_x][i_y][i_z] = 1 # the bin has at least one point (bin is not empty)
voxel_xyz[i_x, i_y, i_z, :] = point_grid[i_idx, :]
voxel_rgb[i_x, i_y, i_z, :] = rgb[i_idx, :]
# voxel_xyzrgb[i_x, i_y, i_z, :3] = point_grid[i_idx, :]
# voxel_xyzrgb[i_x, i_y, i_z, 3:] = rgb[i_idx, :]
# output --- 3D Tensor, 240 x 144 x 240
del depth, gx, gy, pt_cam, pt_world, pt_world2, point_grid # Release Memory
# voxel_binary = voxel_binary.reshape((1,) + voxel_binary.shape) # channel first, (1, W, H, D)
# voxel_binary = voxel_binary.reshape(voxel_binary.shape + (1,)) # channel last , (W, H, D, 1)
return voxel_binary, voxel_rgb, voxel_xyz # voxel_xyzrgb # (W, H, D), (W, H, D, 3), (W, H, D, 6)
@classmethod
def _depth2ply(cls, depth, ply_filename):
# depth = cls._read_depth(depth_filename)
# ---- Get point in camera coordinate
gx, gy = np.meshgrid(range(W), range(H))
point_cam = np.zeros((H, W, 3), dtype=float)
point_cam[:, :, 0] = (gx - cam_K[0][2]) * depth / cam_K[0][0] # x
point_cam[:, :, 1] = (gy - cam_K[1][2]) * depth / cam_K[1][1] # y
point_cam[:, :, 2] = depth # z, in meter
ply_head = 'ply\n' \
'format ascii 1.0\n' \
'element vertex %d\n' \
'property float x\n' \
'property float y\n' \
'property float z\n' \
'end_header' % depth.size
# ply_filename = depth_filename[:-4] + '_depth2ply.ply'
ply_data = point_cam.reshape((-1, 3))
np.savetxt(ply_filename, ply_data, fmt="%f %f %f", header=ply_head, comments='')
print('Saved-->{}'.format(ply_filename))
@classmethod
def _depth_voxel2ply(cls, voxel_val, ply_filename, encoding='BINARY'):
"""ply: x y z of voxels from depth. only save the voxels containing points"""
STSDF_EMPTY = np.float32(0) # 1 is surface, -2 is occluded, 0 is free
TSDF_EMPTY = np.float32(0.001) # -0.001 is surface, -0.001 is occluded, 0.001 is free
# ---- get size
size = voxel_val.shape
# print('size', size)
_x, _y, _z = cls._get_xyz(size)
_x = _x.flatten()
_y = _y.flatten()
_z = _z.flatten()
ply_data_grid = zip(_x, _y, _z, voxel_val.flatten())
ply_data = []
if encoding == 'BINARY': # TODO RGB XYZ XYZRGB ...
for i_idx in range(len(ply_data_grid)):
if ply_data_grid[i_idx][3] > 0: # 0 is empty
ply_data.append(ply_data_grid[i_idx])
if encoding == 'STSDF':
for i_idx in range(len(ply_data_grid)):
# if ply_data_grid[i_idx][3] < 1: # 0 is surface, -1 is occluded
# if ply_data_grid[i_idx][3] != STSDF_EMPTY: # 1 is surface, -1 is occluded, 0 is free
# ply_data.append(ply_data_grid[i_idx])
if ply_data_grid[i_idx][3] == 1: # 1 is surface
ply_data.append(ply_data_grid[i_idx])
if encoding == 'TSDF':
for i_idx in range(len(ply_data_grid)):
# if ply_data_grid[i_idx][3] < -0.0011: # 0 is surface, <0 is occluded, > 0 is free
if ply_data_grid[i_idx][3] != TSDF_EMPTY:
ply_data.append(ply_data_grid[i_idx])
if encoding == 'EMPTY':
for i_idx in range(len(ply_data_grid)):
# if ply_data_grid[i_idx][3] > 0: # 0 is free
if ply_data_grid[i_idx][3] != 0: # 0 is free
ply_data.append(ply_data_grid[i_idx])
if len(ply_data) == 0:
print('From _depth_voxel2ply(): NO valid data. {}'.format(ply_filename))
return
ply_head = 'ply\n' \
'format ascii 1.0\n' \
'element vertex %d\n' \
'property float x\n' \
'property float y\n' \
'property float z\n' \
'property int label\n' \
'end_header' % len(ply_data)
np.savetxt(ply_filename, ply_data, fmt="%d %d %d %d", header=ply_head, comments='')
print('Saved-->{}'.format(ply_filename))
@classmethod
def _gn_voxel2ply(cls, voxel_val, ply_filename, encoding='GN'):
# ---- get size
size = voxel_val.shape
# print('size', size)
_x, _y, _z = cls._get_xyz(size)
_x = _x.flatten()
_y = _y.flatten()
_z = _z.flatten()
# g = np.zeros(voxel_val.size, dtype=np.float32)
# b = np.zeros(voxel_val.size, dtype=np.float32)
gb = np.zeros(voxel_val.size, dtype=np.uint8)
gn_voxel_val = (voxel_val * 255/1.8).astype(np.uint8) # max(gn) = 1.8
ply_data_grid = zip(_x, _y, _z, gn_voxel_val.flatten(), gb, gb)
ply_data = []
if encoding == 'GN':
for i_idx in range(len(ply_data_grid)):
if ply_data_grid[i_idx][3] > 0: # 0 is free
ply_data.append(ply_data_grid[i_idx])
if len(ply_data) == 0:
print('From _depth_voxel2ply(): NO valid data. {}'.format(ply_filename))
return
ply_head = 'ply\n' \
'format ascii 1.0\n' \
'element vertex %d\n' \
'property float x\n' \
'property float y\n' \
'property float z\n' \
'property uchar red\n' \
'property uchar green\n' \
'property uchar blue\n' \
'end_header' % len(ply_data)
np.savetxt(ply_filename, ply_data, fmt="%d %d %d %d %d %d", header=ply_head, comments='')
print('Saved-->{}'.format(ply_filename))
@classmethod
def labeled_voxel2ply(cls, vox_labeled, ply_filename): # TODO: 该函数存在内存泄露风险
"""Save labeled voxels to disk in colored-point cloud format: x y z r g b, with '.ply' suffix
注意 vox_labeled.shape: (W, H, D)
""" #
# ---- Check data type, numpy ndarray
if type(vox_labeled) is not np.ndarray:
raise Exception("Oops! Type of vox_labeled should be 'numpy.ndarray', not {}.".format(type(vox_labeled)))
# ---- Check data validation
if np.amax(vox_labeled) == 0:
print('Oops! All voxel is labeled empty.')
return
# ---- get size
size = vox_labeled.shape
# ---- Convert to list
vox_labeled = vox_labeled.flatten()
# ---- Get X Y Z
_x, _y, _z = cls._get_xyz(size)
_x = _x.flatten()
_y = _y.flatten()
_z = _z.flatten()
# ---- Get R G B
vox_labeled[vox_labeled == 255] = 0 # empty
# vox_labeled[vox_labeled == 255] = 12 # ignore
_rgb = colorMap[vox_labeled[:]]
# ---- Get X Y Z R G B
xyz_rgb = zip(_x, _y, _z, _rgb[:, 0], _rgb[:, 1], _rgb[:, 2])
# xyz_rgb = zip(_z, _y, _x, _rgb[:, 0], _rgb[:, 1], _rgb[:, 2]) # 将X轴和Z轴交换，用于meshlab显示
# ---- Get ply data without empty voxel
xyz_rgb = np.array(xyz_rgb)
ply_data = xyz_rgb[np.where(vox_labeled > 0)]
if len(ply_data) == 0:
raise Exception("Oops! That was no valid ply data.")
ply_head = 'ply\n' \
'format ascii 1.0\n' \
'element vertex %d\n' \
'property float x\n' \
'property float y\n' \
'property float z\n' \
'property uchar red\n' \
'property uchar green\n' \
'property uchar blue\n' \
'end_header' % len(ply_data)
# ---- Save ply data to disk
np.savetxt(ply_filename, ply_data, fmt="%d %d %d %d %d %d", header=ply_head, comments='') # It takes 20s
del vox_labeled, _x, _y, _z, _rgb, xyz_rgb, ply_data, ply_head
# print('Saved-->{}'.format(ply_filename))
@staticmethod
def _downsample_label(label, downscale=4):
r"""downsample the labeled data, new version takes only about 0.6s on CPU
Shape:
label, (240, 144, 240)
label_downscale, if downsample==4, then (60, 36, 60)
"""
if downscale == 1:
return label
ds = int(downscale)
small_size = (int(voxel_SIZE[0] / ds), int(voxel_SIZE[1] / ds), int(voxel_SIZE[2] / ds)) # small size
label_downscale = np.zeros(small_size, dtype=np.uint8)
empty_t = 0.95 * ds * ds * ds # threshold
s01 = small_size[0] * small_size[1]
label_i = np.zeros((ds, ds, ds), dtype=np.int32)
for i in range(small_size[0]*small_size[1]*small_size[2]):
z = i / s01
y = (i - z * s01) / small_size[0]
x = i - z * s01 - y * small_size[0]
x = int(x)
y = int(y)
z = int(z)
label_i[:, :, :] = label[x * ds:(x + 1) * ds, y * ds:(y + 1) * ds, z * ds:(z + 1) * ds]
label_bin = label_i.flatten()
# zero_count = np.array(np.where(np.logical_or(label_i < 0.001, label_i > 254))).size
# zero_count = np.array(np.where(np.logical_or(label_bin == 0, label_bin == 255))).size
# if zero_count < empty_t:
# label_i_s = label_bin[np.nonzero(label_bin)] # get the none empty class labels
# label_downscale[x, y, z] = np.argmax(np.bincount(label_i_s))
zero_count_0 = np.array(np.where(label_bin == 0)).size
zero_count_255 = np.array(np.where(label_bin == 255)).size
zero_count = zero_count_0 + zero_count_255
if zero_count > empty_t:
label_downscale[x, y, z] = 0 if zero_count_0 > zero_count_255 else 255
else:
# label_i_s = label_bin[np.nonzero(label_bin)] # get the none empty class labels
label_i_s = label_bin[np.where(np.logical_and(label_bin > 0, label_bin < 255))]
label_downscale[x, y, z] = np.argmax(np.bincount(label_i_s))
return label_downscale
@staticmethod # 0 for empty, 1 for surface, -1 for occ
def _downsample_stsdf(stsdf, downscale=4):
r"""
Shape:
stsdf, (240, 144, 240)
stsdf_downscale, (60, 36, 60)
"""
if downscale == 1:
return stsdf
STSDF_EMPTY = 0
STSDF_SURFACE = 1
STSDF_OCCLUD = -1
ds = downscale
small_size = (np.int32(stsdf.shape[0] / ds), np.int32(stsdf.shape[1] / ds), np.int32(stsdf.shape[2] / ds))
t = ds * ds * ds * 0.05 # t = 0
stsdf_downscale = np.zeros(small_size, dtype=np.float32) # init 0 for empty
# stsdf_downscale = np.full(small_size, -2.0, dtype=np.float32) # init -2 for occ
s01 = small_size[0] * small_size[1]
stsdf_i = np.ones((ds, ds, ds), dtype=np.float32)
for i in range(small_size[0] * small_size[1] * small_size[2]):
z = int(i / s01)
y = int((i - z * s01) / small_size[0])
x = int(i - z * s01 - y * small_size[0])
stsdf_i[:, :, :] = stsdf[x * ds:(x + 1) * ds, y * ds:(y + 1) * ds, z * ds:(z + 1) * ds]
stsdf_bin = stsdf_i.flatten()
none_empty_count = np.array(np.where(stsdf_bin != STSDF_EMPTY)).size # 0 for empty
# if none_empty_count > t:
# # surface_count = np.array(np.where(stsdf_bin == 1)).size
# # occluded_count = np.array(np.where(stsdf_bin == -2)).size
# surface_count = np.array(np.where(stsdf_bin > 0)).size
# occluded_count = np.array(np.where(stsdf_bin < 0)).size
# stsdf_downscale[x, y, z] = STSDF_SURFACE if surface_count >= occluded_count else STSDF_OCCLUD
if none_empty_count > t:
surface_count = np.array(np.where(stsdf_bin > 0)).size
# occluded_count = np.array(np.where(stsdf_bin < 0)).size
# stsdf_downscale[x, y, z] = STSDF_SURFACE if surface_count > t else STSDF_OCCLUD
# stsdf_downscale[x, y, z] = STSDF_SURFACE if (surface_count > t or surface_count >= occluded_count) else STSDF_OCCLUD
stsdf_downscale[x, y, z] = STSDF_SURFACE if surface_count > 2 else STSDF_OCCLUD # 至少3个点才视为surface
# else:
# stsdf_downscale[x, y, z] = 0 # 0 is empty, default is 0
return stsdf_downscale
@staticmethod
def _downsample_tsdf(tsdf, downscale=4): # 仅在Get None empty 时会用到
r"""
Shape:
tsdf, (240, 144, 240)
tsdf_downscale, (60, 36, 60), (stsdf.shape[0]/4, stsdf.shape[1]/4, stsdf.shape[2]/4)
"""
if downscale == 1:
return tsdf
# TSDF_EMPTY = np.float32(0.001)
# TSDF_SURFACE: 1, sign >= 0
# TSDF_OCCLUD: sign < 0 np.float32(-0.001)
ds = downscale
small_size = (int(tsdf.shape[0] / ds), int(tsdf.shape[1] / ds), int(tsdf.shape[2] / ds))
tsdf_downscale = np.ones(small_size, dtype=np.float32) * np.float32(0.001) # init 0.001 for empty
s01 = small_size[0] * small_size[1]
tsdf_sr = np.ones((ds, ds, ds), dtype=np.float32) # search region
for i in range(small_size[0] * small_size[1] * small_size[2]):
z = int(i / s01)
y = int((i - z * s01) / small_size[0])
x = int(i - z * s01 - y * small_size[0])
tsdf_sr[:, :, :] = tsdf[x * ds:(x + 1) * ds, y * ds:(y + 1) * ds, z * ds:(z + 1) * ds]
tsdf_bin = tsdf_sr.flatten()
# none_empty_count = np.array(np.where(tsdf_bin != TSDF_EMPTY)).size
none_empty_count = np.array(np.where(np.logical_or(tsdf_bin <= 0, tsdf_bin == 1))).size
if none_empty_count > 0:
# surface_count = np.array(np.where(stsdf_bin == 1)).size
# occluded_count = np.array(np.where(stsdf_bin == -2)).size
# surface_count = np.array(np.where(tsdf_bin > 0)).size # 这个存在问题
surface_count = np.array(np.where(tsdf_bin == 1)).size
# occluded_count = np.array(np.where(tsdf_bin < 0)).size
# tsdf_downscale[x, y, z] = 0 if surface_count > occluded_count else np.float32(-0.001)
tsdf_downscale[x, y, z] = 1 if surface_count > 2 else np.float32(-0.001) # 1 or 0 ?
# else:
# tsdf_downscale[x, y, z] = empty # TODO 不应该将所有值均设为0.001
return tsdf_downscale
@staticmethod
def get_nonempty(voxels, encoding): # Get none empty from depth voxels
data = np.zeros(voxels.shape, dtype=np.float32) # init 0 for empty
# if encoding == 'STSDF': # surface, empty, occulted: 1, 0, -1
# data[voxels == 1] = 1
# return data
if encoding == 'STSDF': # surface, empty, occulted: 1, 0, -1
data[voxels != 0] = 1
surface = np.array(np.where(voxels == 1)) # surface=1
elif encoding == 'TSDF':
data[np.where(np.logical_or(voxels <= 0, voxels == 1))] = 1
surface = np.array(np.where(voxels == 1)) # surface
# surface = np.array(np.where(np.logical_and(voxels > 0, voxels != np.float32(0.001)))) # surface
else:
raise Exception("Encoding error: {} is not validate".format(encoding))
min_idx = np.amin(surface, axis=1)
max_idx = np.amax(surface, axis=1)
# print('min_idx, max_idx', min_idx, max_idx)
# data[:a], data[a]不包含在内, data[b:], data[b]包含在内
# min_idx = min_idx
max_idx = max_idx + 1
# 本该扩大一圈就够了，但由于GT标注的不是很精确，故在高分辨率情况下，多加大一圈
# min_idx = min_idx - 1
# max_idx = max_idx + 2
min_idx[min_idx < 0] = 0
max_idx[0] = min(voxels.shape[0], max_idx[0])
max_idx[1] = min(voxels.shape[1], max_idx[1])