forked from art-programmer/PlaneNet
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_planenet.py
1326 lines (1093 loc) · 71.5 KB
/
train_planenet.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
import tensorflow as tf
import numpy as np
np.set_printoptions(precision=2, linewidth=200)
import cv2
import os
import time
import sys
import argparse
import glob
import PIL
import scipy.ndimage as ndimage
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from utils import *
from modules import *
from planenet import PlaneNet
from RecordReaderAll import *
from crfasrnn.crfasrnn_layer import CrfRnnLayer
#from SegmentationRefinement import *
#training_flag: toggle dropout and batch normalization mode
#it's true for training and false for validation, testing, prediction
#it also controls which data batch to use (*_train or *_val)
def build_graph(img_inp_train, img_inp_val, training_flag, options):
with tf.device('/gpu:%d'%options.gpu_id):
img_inp = tf.cond(training_flag, lambda: img_inp_train, lambda: img_inp_val)
net = PlaneNet({'img_inp': img_inp}, is_training=training_flag, options=options)
#global predictions
plane_pred = net.layers['plane_pred']
segmentation_pred = net.layers['segmentation_pred']
non_plane_mask_pred = net.layers['non_plane_mask_pred']
non_plane_depth_pred = net.layers['non_plane_depth_pred']
non_plane_normal_pred = net.layers['non_plane_normal_pred']
non_plane_normal_pred = tf.nn.l2_normalize(non_plane_normal_pred, dim=-1)
if False:
plane_pred = gt_dict['plane']
non_plane_mask_pred = gt_dict['non_plane_mask'] * 10
non_plane_depth_pred = gt_dict['depth']
non_plane_normal_pred = gt_dict['normal']
segmentation_pred = gt_dict['segmentation'][:, :, :, :options.numOutputPlanes] * 10
pass
if abs(options.crfrnn) > 0:
with tf.variable_scope('crfrnn'):
all_segmentations = crfrnnModule([tf.concat([segmentation_pred, non_plane_mask_pred], axis=3), img_inp * 255], image_dims=(HEIGHT, WIDTH), num_classes=options.numOutputPlanes + 1, theta_alpha=30, theta_beta=10, theta_gamma=1, num_iterations=abs(options.crfrnn))
#all_segmentations = CrfRnnLayer(image_dims=(HEIGHT, WIDTH), num_classes=options.numOutputPlanes + 1, theta_alpha=80., theta_beta=3., theta_gamma=3., num_iterations=abs(options.crfrnn), name='crfrnn')([tf.concat([segmentation_pred, non_plane_mask_pred], axis=3), img_inp * 255])
segmentation_pred = all_segmentations[:, :, :, :options.numOutputPlanes]
non_plane_mask_pred = all_segmentations[:, :, :, options.numOutputPlanes:]
pass
pass
global_pred_dict = {'plane': plane_pred, 'segmentation': segmentation_pred, 'non_plane_mask': non_plane_mask_pred, 'non_plane_depth': non_plane_depth_pred, 'non_plane_normal': non_plane_normal_pred}
if options.predictBoundary:
global_pred_dict['boundary'] = net.layers['boundary_pred']
else:
global_pred_dict['boundary'] = tf.zeros((options.batchSize, HEIGHT, WIDTH, 2))
pass
if options.predictConfidence:
global_pred_dict['confidence'] = net.layers['plane_confidence_pred']
pass
if options.predictSemantics:
global_pred_dict['semantics'] = net.layers['semantics_pred']
pass
#local predictions
if options.predictLocal:
local_pred_dict = {'score': net.layers['local_score_pred'], 'plane': net.layers['local_plane_pred'], 'mask': net.layers['local_mask_pred']}
else:
local_pred_dict = {}
pass
#deep supervision
deep_pred_dicts = []
for layer in options.deepSupervisionLayers:
pred_dict = {'plane': net.layers[layer+'_plane_pred'], 'segmentation': net.layers[layer+'_segmentation_pred'], 'non_plane_mask': net.layers[layer+'_non_plane_mask_pred']}
#if options.predictConfidence:
#pred_dict['confidence'] = net.layers[layer+'_plane_confidence_pred']
#pass
deep_pred_dicts.append(pred_dict)
continue
if options.anchorPlanes:
anchors_np = np.load('dump/anchors_' + options.hybrid + '.npy')
anchors = tf.reshape(tf.constant(anchors_np.reshape(-1)), anchors_np.shape)
anchors = tf.tile(tf.expand_dims(anchors, 0), [options.batchSize, 1, 1])
all_pred_dicts = deep_pred_dicts + [global_pred_dict]
for pred_index, pred_dict in enumerate(all_pred_dicts):
all_pred_dicts[pred_index]['plane'] += anchors
continue
pass
pass
return global_pred_dict, local_pred_dict, deep_pred_dicts
def build_loss(img_inp_train, img_inp_val, global_pred_dict, deep_pred_dicts, global_gt_dict_train, global_gt_dict_val, training_flag, options):
from nndistance import tf_nndistance
with tf.device('/gpu:%d'%options.gpu_id):
debug_dict = {}
img_inp = tf.cond(training_flag, lambda: img_inp_train, lambda: img_inp_val)
global_gt_dict = {}
for name in global_gt_dict_train.keys():
global_gt_dict[name] = tf.cond(training_flag, lambda: global_gt_dict_train[name], lambda: global_gt_dict_val[name])
continue
# local_gt_dict = {}
# for name in local_gt_dict_train.keys():
# local_gt_dict[name] = tf.cond(tf.equal(training_flag % 2, 0), lambda: local_gt_dict_train[name], lambda: local_gt_dict_val[name])
# continue
plane_parameters = tf.reshape(global_pred_dict['plane'], (-1, 3))
info = global_gt_dict['info'][0]
plane_depths = planeDepthsModule(plane_parameters, WIDTH, HEIGHT, info)
plane_depths = tf.transpose(tf.reshape(plane_depths, [HEIGHT, WIDTH, -1, options.numOutputPlanes]), [2, 0, 1, 3])
non_plane_depth = global_pred_dict['non_plane_depth']
all_depths = tf.concat([plane_depths, non_plane_depth], axis=3)
validPlaneMask = tf.cast(tf.less(tf.tile(tf.expand_dims(tf.range(options.numOutputPlanes), 0), [options.batchSize, 1]), tf.expand_dims(global_gt_dict['num_planes'], -1)), tf.float32)
backwardLossWeight = options.backwardLossWeight
#plane loss and segmentation loss (summation over deep supervisions and final supervision)
all_pred_dicts = deep_pred_dicts + [global_pred_dict]
plane_loss = tf.constant(0.0)
segmentation_loss = tf.constant(0.0)
plane_confidence_loss = tf.constant(0.0)
diverse_loss = tf.constant(0.0)
#keep forward map (segmentation gt) from previous supervision so that we can have same matching for all supervisions (options.sameMatching = 1)
previous_plane_gt = None
previous_plane_confidence_gt = None
previous_segmentation_gt = None
if options.anchorPlanes:
anchors_np = np.load('dump/anchors_' + options.hybrid + '.npy')
anchors = tf.reshape(tf.constant(anchors_np.reshape(-1)), anchors_np.shape)
anchors = tf.tile(tf.expand_dims(anchors, 0), [options.batchSize, 1, 1])
dists_forward, map_forward, dists_backward, _ = tf_nndistance.nn_distance(global_gt_dict['plane'], anchors)
forward_map = tf.one_hot(map_forward, depth=options.numOutputPlanes, axis=-1)
forward_map *= tf.expand_dims(validPlaneMask, -1)
#number of ground truth mapped for each prediction
num_matches = tf.transpose(tf.reduce_sum(forward_map, axis=1, keep_dims=True), [0, 2, 1])
previous_plane_gt = tf.transpose(tf.matmul(global_gt_dict['plane'], forward_map, transpose_a=True), [0, 2, 1]) / tf.maximum(num_matches, 1e-4)
previous_plane_confidence_gt = tf.cast(num_matches > 0.5, tf.float32)
segmentation_gt_shuffled = tf.reshape(tf.matmul(tf.reshape(global_gt_dict['segmentation'], [-1, HEIGHT * WIDTH, options.numOutputPlanes]), forward_map), [-1, HEIGHT, WIDTH, options.numOutputPlanes])
segmentation_gt_shuffled = tf.concat([segmentation_gt_shuffled, global_gt_dict['non_plane_mask']], axis=3)
previous_segmentation_gt = segmentation_gt_shuffled
pass
for pred_index, pred_dict in enumerate(all_pred_dicts):
if (options.sameMatching and pred_index > 0) or options.anchorPlanes:
#use matching from previous supervision and map ground truth planes based on the mapping
plane_loss += tf.reduce_mean(tf.squared_difference(pred_dict['plane'], previous_plane_gt) * previous_plane_confidence_gt) * 10000
#all segmentations is the concatenation of plane segmentations and non plane mask
all_segmentations = tf.concat([pred_dict['segmentation'], pred_dict['non_plane_mask']], axis=3)
if options.crf > 0:
all_segmentations_softmax = tf.nn.softmax(all_segmentations)
with tf.variable_scope('crf'):
planesY = global_pred_dict['plane'][:, :, 1]
planesD = tf.maximum(tf.norm(global_pred_dict['plane'], axis=-1), 1e-4)
planesY /= planesD
planesY = tf.concat([planesY, tf.ones((options.batchSize, 1)) * 100], axis=1)
imageDiff = calcImageDiff(img_inp)
all_segmentations_softmax, _ = segmentationRefinementModule(all_segmentations_softmax, all_depths, planesY, imageDiff, numIterations=options.crf, numOutputPlanes=21)
segmentation_loss += tf.reduce_mean(-tf.reduce_sum(previous_segmentation_gt* tf.log(tf.maximum(all_segmentations_softmax, 1e-31)), axis=-1)) * 1000
pass
pass
else:
segmentation_loss += tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=all_segmentations, labels=previous_segmentation_gt)) * 1000
pass
else:
#calculate new matching by finding nearest neighbors again
dists_forward, map_forward, dists_backward, _ = tf_nndistance.nn_distance(global_gt_dict['plane'], pred_dict['plane'])
dists_forward *= validPlaneMask
dists_forward = tf.reduce_mean(dists_forward)
dists_backward = tf.reduce_mean(dists_backward)
plane_loss += (dists_forward + dists_backward * backwardLossWeight) * 10000
forward_map = tf.one_hot(map_forward, depth=options.numOutputPlanes, axis=-1)
forward_map *= tf.expand_dims(validPlaneMask, -1)
#number of ground truth mapped for each prediction
num_matches = tf.transpose(tf.reduce_sum(forward_map, axis=1, keep_dims=True), [0, 2, 1])
previous_plane_gt = tf.transpose(tf.matmul(global_gt_dict['plane'], forward_map, transpose_a=True), [0, 2, 1]) / tf.maximum(num_matches, 1e-4)
previous_plane_confidence_gt = tf.cast(num_matches > 0.5, tf.float32)
segmentation_gt_shuffled = tf.reshape(tf.matmul(tf.reshape(global_gt_dict['segmentation'], [-1, HEIGHT * WIDTH, options.numOutputPlanes]), forward_map), [-1, HEIGHT, WIDTH, options.numOutputPlanes])
segmentation_gt_shuffled = tf.concat([segmentation_gt_shuffled, global_gt_dict['non_plane_mask']], axis=3)
previous_segmentation_gt = segmentation_gt_shuffled
all_segmentations = tf.concat([pred_dict['segmentation'], pred_dict['non_plane_mask']], axis=3)
segmentation_loss += tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=all_segmentations, labels=segmentation_gt_shuffled)) * 1000
debug_dict['segmentation'] = segmentation_gt_shuffled
pass
if options.diverseLoss:
plane_diff = tf.reduce_sum(tf.pow(tf.expand_dims(pred_dict['plane'], 1) - tf.expand_dims(pred_dict['plane'], 2), 2), axis=3)
plane_diff = tf.matrix_set_diag(plane_diff, tf.ones((options.batchSize, options.numOutputPlanes)))
minPlaneDiff = 0.5
diverse_loss += tf.reduce_mean(tf.clip_by_value(1 - plane_diff / minPlaneDiff, 0, 1)) * 10000
pass
continue
if options.crf == 0:
all_segmentations_softmax = tf.nn.softmax(all_segmentations)
pass
#depth loss
validDepthMask = tf.cast(tf.greater(global_gt_dict['depth'], 1e-4), tf.float32)
depth_loss = tf.constant(0.0)
if options.depthLoss == 1:
depth_loss += tf.reduce_mean(tf.reduce_sum(tf.squared_difference(all_depths, global_gt_dict['depth']) * all_segmentations_softmax, axis=3, keep_dims=True) * validDepthMask) * 20000
elif options.depthLoss == 2:
depthDiff = tf.abs(all_depths - global_gt_dict['depth'])
c = 0.3
absMask = tf.cast(tf.less(depthDiff, c), tf.float32)
depthDiff = depthDiff * absMask + (tf.pow(depthDiff, 2) + tf.pow(c, 2)) / (2 * c) * (1 - absMask)
depth_loss += tf.reduce_mean(tf.reduce_sum(depthDiff * all_segmentations_softmax, axis=3, keep_dims=True) * validDepthMask) * 10000
elif options.depthLoss == 3:
depth_softmax = tf.reduce_sum(all_depths * all_segmentations_softmax, axis=3, keep_dims=True)
depthDiff = tf.abs(depth_softmax - global_gt_dict['depth'])
c = 0.3
absMask = tf.cast(tf.less(depthDiff, c), tf.float32)
depthDiff = depthDiff * absMask + (tf.pow(depthDiff, 2) + tf.pow(c, 2)) / (2 * c) * (1 - absMask)
depth_loss += tf.reduce_mean(depthDiff * validDepthMask) * 10000
elif options.depthLoss == 4:
S = tf.one_hot(tf.argmax(all_segmentations, 3), depth=options.numOutputPlanes + 1)
depth_one_hot = tf.reduce_sum(all_depths * S, axis=3, keep_dims=True)
depthDiff = tf.abs(depth_one_hot - global_gt_dict['depth'])
c = 0.3
absMask = tf.cast(tf.less(depthDiff, c), tf.float32)
depthDiff = depthDiff * absMask + (tf.pow(depthDiff, 2) + tf.pow(c, 2)) / (2 * c) * (1 - absMask)
depth_loss += tf.reduce_mean(depthDiff * validDepthMask) * 10000
pass
if options.predictPixelwise == 1:
depth_diff = global_pred_dict['non_plane_depth'] - global_gt_dict['depth']
depth_diff_gx = depth_diff - tf.concat([tf.ones([options.batchSize, HEIGHT, 1, 1]), depth_diff[:, :, :WIDTH - 1]], axis=2)
depth_diff_gy = depth_diff - tf.concat([tf.ones([options.batchSize, 1, WIDTH, 1]), depth_diff[:, :HEIGHT - 1]], axis=1)
numValidPixels = tf.reduce_sum(validDepthMask, axis=[1, 2, 3])
depth_loss += tf.reduce_mean(tf.reduce_sum(tf.pow(depth_diff * validDepthMask, 2), axis=[1, 2, 3]) / numValidPixels - 0.5 * tf.pow(tf.reduce_sum(depth_diff * validDepthMask, axis=[1, 2, 3]) / numValidPixels, 2) + tf.reduce_sum((tf.pow(depth_diff_gx, 2) + tf.pow(depth_diff_gy, 2)) * validDepthMask, axis=[1, 2, 3]) / numValidPixels) * 1000
#depth_loss += tf.reduce_mean(tf.squared_difference(global_pred_dict['non_plane_depth'], global_gt_dict['depth']) * validDepthMask) * 10000
valid_normal_mask = tf.squeeze(tf.cast(tf.less(tf.slice(global_gt_dict['info'], [0, 19], [options.batchSize, 1]), 2), tf.float32))
normal_gt = tf.nn.l2_normalize(global_gt_dict['normal'], dim=-1)
normal_loss = tf.reduce_mean(tf.reduce_sum(-global_pred_dict['non_plane_normal'] * normal_gt * validDepthMask, axis=[1, 2, 3]) / numValidPixels * valid_normal_mask) * 1000
#normal_loss = tf.reduce_mean(tf.reduce_mean(tf.squared_difference(global_pred_dict['non_plane_normal'], global_gt_dict['normal']) * validDepthMask, axis=[1, 2, 3]) * valid_normal_mask) * 1000
else:
normal_loss = tf.constant(0.0)
pass
if options.predictSemantics:
valid_semantics_mask = tf.squeeze(tf.cast(tf.not_equal(tf.slice(global_gt_dict['info'], [0, 19], [options.batchSize, 1]), 1), tf.float32))
semantics_loss = tf.reduce_mean(tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=global_pred_dict['semantics'], labels=global_gt_dict['semantics']), axis=[1, 2]) * valid_semantics_mask) * 1000
else:
semantics_loss = tf.constant(0.0)
pass
local_score_loss = tf.constant(0.0)
local_plane_loss = tf.constant(0.0)
local_mask_loss = tf.constant(0.0)
#boundary loss
boundary_loss = tf.constant(0.0)
if False:
kernel_size = 3
padding = (kernel_size - 1) / 2
neighbor_kernel_array = gaussian(kernel_size, kernel_size)
neighbor_kernel_array[(kernel_size - 1) / 2][(kernel_size - 1) / 2] = 0
neighbor_kernel_array /= neighbor_kernel_array.sum()
neighbor_kernel_array *= -1
neighbor_kernel_array[(kernel_size - 1) / 2][(kernel_size - 1) / 2] = 1
neighbor_kernel = tf.constant(neighbor_kernel_array.reshape(-1), shape=neighbor_kernel_array.shape, dtype=tf.float32)
neighbor_kernel = tf.reshape(neighbor_kernel, [kernel_size, kernel_size, 1, 1])
#calculate boundary ground truth on-the-fly as the calculation is subject to change
depth_diff = tf.abs(tf.nn.depthwise_conv2d(global_gt_dict['depth'], neighbor_kernel, strides=[1, 1, 1, 1], padding='VALID'))
depth_diff = tf.pad(depth_diff, paddings = [[0, 0], [padding, padding], [padding, padding], [0, 0]])
max_depth_diff = 0.1
depth_boundary = tf.greater(depth_diff, max_depth_diff)
plane_region = tf.nn.max_pool(1 - global_gt_dict['non_plane_mask'], ksize=[1, kernel_size, kernel_size, 1], strides=[1, 1, 1, 1], padding='SAME', name='max_pool')
segmentation_eroded = 1 - tf.nn.max_pool(1 - global_gt_dict['segmentation'], ksize=[1, 3, 3, 1], strides=[1, 1, 1, 1], padding='SAME', name='max_pool')
plane_region -= tf.reduce_max(segmentation_eroded, axis=3, keep_dims=True)
boundary = tf.cast(depth_boundary, tf.float32) * plane_region
smooth_boundary = global_gt_dict['smooth_boundary']
boundary_gt = tf.concat([smooth_boundary, tf.maximum(boundary - smooth_boundary, 0)], axis=3)
pass
if options.boundaryLoss == 1 and False:
all_segmentations_pred = all_segmentations_softmax
all_segmentations_min = 1 - tf.nn.max_pool(1 - all_segmentations_pred, ksize=[1, 3, 3, 1], strides=[1, 1, 1, 1], padding='SAME')
segmentation_diff = tf.reduce_max(all_segmentations_pred - all_segmentations_min, axis=3, keep_dims=True)
depth_pred = tf.reduce_sum(tf.multiply(all_depths, all_segmentations_pred), axis=3, keep_dims=True)
depth_neighbor = tf.nn.depthwise_conv2d(depth_pred, neighbor_kernel, strides=[1, 1, 1, 1], padding='SAME')
minDepthDiff = 0.02
depth_diff = tf.clip_by_value(tf.squared_difference(depth_pred, depth_neighbor) - pow(minDepthDiff, 2), 0, 1)
boundary = tf.reduce_max(global_gt_dict['boundary'], axis=3, keep_dims=True)
smooth_boundary = tf.slice(boundary_gt, [0, 0, 0, 0], [options.batchSize, HEIGHT, WIDTH, 1])
smooth_mask = segmentation_diff + boundary - 2 * segmentation_diff * boundary + depth_diff * smooth_boundary
margin = 0.0
smooth_mask = tf.nn.relu(smooth_mask - margin)
#boundary_loss += tf.reduce_mean(smooth_mask * plane_mask) * 1000
boundary_loss += tf.reduce_mean(smooth_mask) * 1000
elif options.boundaryLoss == 3:
S = tf.one_hot(tf.argmax(all_segmentations, 3), depth=options.numOutputPlanes + 1)
#sigmaDepthDiff = 0.5
#DS_diff = tf.exp(-tf.pow(1 - tf.clip_by_value(tf.abs(all_depths - tf.reduce_sum(all_depths * S, 3, keep_dims=True)), 0, 1), 2) / sigmaDepthDiff) - tf.exp(-1 / sigmaDepthDiff) * S
planesY = tf.slice(global_pred_dict['plane'], [0, 0, 1], [options.batchSize, options.numOutputPlanes, 1])
planesD = tf.maximum(tf.norm(global_pred_dict['plane'], axis=-1, keep_dims=True), 1e-4)
planesY /= planesD
#normalY = tf.reduce_sum(tf.slice(S, [0, 0, 0, 0], [options.batchSize, HEIGHT, WIDTH, options.numOutputPlanes]) * tf.reshape(planesY, [options.batchSize, 1, 1, options.numOutputPlanes]), axis=3, keep_dims=True)
planesY = tf.concat([planesY, tf.ones((options.batchSize, 1, 1)) * 100], axis=1)
normalY = tf.reduce_sum(S * tf.reshape(planesY, [options.batchSize, 1, 1, -1]), axis=3, keep_dims=True)
maxDepthDiff = 0.1
labelDiffWeight = 0.05
depth_one_hot = tf.reduce_sum(all_depths * S, 3, keep_dims=True)
depth_diff = tf.clip_by_value(tf.pow((plane_depths - depth_one_hot) * normalY / maxDepthDiff, 2), 0, 1)
#depth_diff *= normalY
depth_diff = tf.concat([depth_diff, 1 - tf.slice(S, [0, 0, 0, options.numOutputPlanes], [options.batchSize, HEIGHT, WIDTH, 1])], axis=3)
DS_diff = (1 + labelDiffWeight) - tf.exp(-depth_diff) - S * labelDiffWeight
kernel_size = 3
neighbor_kernel_array = gaussian(kernel_size)
neighbor_kernel_array[(kernel_size - 1) / 2][(kernel_size - 1) / 2] = 0
neighbor_kernel_array /= neighbor_kernel_array.sum()
neighbor_kernel = tf.constant(neighbor_kernel_array.reshape(-1), shape=neighbor_kernel_array.shape, dtype=tf.float32)
neighbor_kernel = tf.reshape(neighbor_kernel, [kernel_size, kernel_size, 1, 1])
DS = tf.nn.depthwise_conv2d(DS_diff, tf.tile(neighbor_kernel, [1, 1, options.numOutputPlanes + 1, 1]), strides=[1, 1, 1, 1], padding='SAME')
boundary_loss += tf.reduce_mean(DS * all_segmentations_softmax) * 100000
debug_dict['cost_mask'] = tf.reduce_sum(DS * all_segmentations_softmax, axis=3)
#debug_dict['cost_mask'] = DS * all_segmentations_softmax
#debug_dict['cost_mask'] = tf.reduce_sum(DS * S, axis=3)
elif options.boundaryLoss == 2:
planesY = global_pred_dict['plane'][:, :, 1]
planesD = tf.maximum(tf.norm(global_pred_dict['plane'], axis=-1), 1e-4)
planesY /= planesD
planesY = tf.concat([planesY, tf.ones((options.batchSize, 1)) * 100], axis=1)
#imageDiff = calcImageDiff(img_inp)
#all_segmentations_softmax, _ = segmentationRefinementModule(all_segmentations_softmax, all_depths, planesY, imageDiff, numIterations=options.crf, numOutputPlanes=21)
messages = calcMessages(all_segmentations_softmax, all_depths, planesY, numOutputPlanes=21)
boundary_loss += tf.reduce_mean(messages * all_segmentations_softmax) * 100000
debug_dict['cost_mask'] = messages
pass
if options.predictBoundary and False:
#we predict boundaries directly for post-processing purpose
boundary_loss += tf.reduce_mean(tf.losses.sigmoid_cross_entropy(logits=global_pred_dict['boundary'], multi_class_labels=global_gt_dict['boundary'], weights=tf.maximum(global_gt_dict['boundary'] * 3, 1))) * 1000
pass
label_loss = tf.constant(0.0)
if options.labelLoss == 1:
#label_loss = tf.reduce_mean(tf.reduce_max(all_segmentations_softmax, axis=[1, 2]) * tf.concat([tf.cast(tf.equal(tf.squeeze(num_matches, axis=2), 0), tf.float32), tf.ones([options.batchSize, 1])], axis=1)) * 1000
#label_loss = tf.reduce_mean(tf.log(1 + tf.reduce_sum(all_segmentations_softmax, axis=[1, 2]))) * 100
segmentations_gt = tf.concat([global_gt_dict['segmentation'], global_gt_dict['non_plane_mask']], axis=3)
label_loss = tf.reduce_mean(tf.maximum(tf.reduce_sum(tf.sqrt(tf.reduce_sum(all_segmentations_softmax, axis=[1, 2])), axis=1) - tf.reduce_sum(tf.sqrt(tf.reduce_sum(segmentations_gt, axis=[1, 2])), axis=1), 0)) * 5
#label_loss = tf.reduce_mean(tf.reduce_max(all_segmentations_softmax, axis=[1, 2])) * 1000
pass
#regularization
l2_losses = tf.add_n([options.l2Weight * tf.nn.l2_loss(v) for v in tf.trainable_variables() if 'weights' in v.name])
if options.planeLoss == 0:
plane_loss = tf.constant(0.0)
pass
loss = plane_loss + segmentation_loss + depth_loss + normal_loss + plane_confidence_loss + diverse_loss + boundary_loss + local_score_loss + local_plane_loss + local_mask_loss + label_loss + semantics_loss + l2_losses
#if options.pixelwiseLoss:
#normal_loss = tf.reduce_mean(tf.squared_difference(global_pred_dict['non_plane_normal'], global_gt_dict['normal'])) * 1000
#depth_loss = tf.reduce_mean(tf.squared_difference(global_pred_dict['non_plane_depth'], global_gt_dict['depth']) * validDepthMask) * 1000
#pass
S = tf.one_hot(tf.argmax(all_segmentations, 3), depth=options.numOutputPlanes + 1)
depth_one_hot = tf.reduce_sum(all_depths * S, 3, keep_dims=True)
debug_dict['depth'] = depth_one_hot
loss_dict = {'plane': plane_loss, 'segmentation': segmentation_loss, 'depth': depth_loss, 'normal': normal_loss, 'boundary': boundary_loss, 'diverse': diverse_loss, 'confidence': plane_confidence_loss, 'local_score': local_score_loss, 'local_plane': local_plane_loss, 'local_mask': local_mask_loss, 'label': label_loss, 'semantics': semantics_loss}
pass
return loss, loss_dict, debug_dict
def main(options):
if not os.path.exists(options.checkpoint_dir):
os.system("mkdir -p %s"%options.checkpoint_dir)
pass
if not os.path.exists(options.test_dir):
os.system("mkdir -p %s"%options.test_dir)
pass
min_after_dequeue = 1000
train_inputs = []
val_inputs = []
if '0' in options.hybrid:
train_inputs.append(options.dataFolder + '/planes_SUNCG_train.tfrecords')
val_inputs.append(options.dataFolder + '/planes_SUNCG_val.tfrecords')
pass
if '1' in options.hybrid:
for _ in xrange(10):
train_inputs.append(options.dataFolder + '/planes_nyu_rgbd_train.tfrecords')
train_inputs.append(options.dataFolder + '/planes_nyu_rgbd_labeled_train.tfrecords')
val_inputs.append(options.dataFolder + '/planes_nyu_rgbd_val.tfrecords')
continue
pass
if '2' in options.hybrid:
train_inputs.append(options.dataFolder + '/planes_matterport_train.tfrecords')
val_inputs.append(options.dataFolder + '/planes_matterport_val.tfrecords')
pass
if '3' in options.hybrid:
train_inputs.append(options.dataFolder + '/planes_scannet_train.tfrecords')
val_inputs.append(options.dataFolder + '/planes_scannet_val.tfrecords')
pass
reader_train = RecordReaderAll()
filename_queue_train = tf.train.string_input_producer(train_inputs, num_epochs=10000)
img_inp_train, global_gt_dict_train, local_gt_dict_train = reader_train.getBatch(filename_queue_train, numOutputPlanes=options.numOutputPlanes, batchSize=options.batchSize, min_after_dequeue=min_after_dequeue, getLocal=True)
reader_val = RecordReaderAll()
filename_queue_val = tf.train.string_input_producer(val_inputs, num_epochs=10000)
img_inp_val, global_gt_dict_val, local_gt_dict_val = reader_val.getBatch(filename_queue_val, numOutputPlanes=options.numOutputPlanes, batchSize=options.batchSize, min_after_dequeue=min_after_dequeue, getLocal=True)
training_flag = tf.placeholder(tf.bool, shape=[], name='training_flag')
#global_pred_dict, local_pred_dict, deep_pred_dicts = build_graph(img_inp_train, img_inp_val, img_inp_rgbd_train, img_inp_rgbd_val, img_inp_3d_train, img_inp_3d_val, training_flag, options)
global_pred_dict, local_pred_dict, deep_pred_dicts = build_graph(img_inp_train, img_inp_val, training_flag, options)
#loss, loss_dict, _ = build_loss(global_pred_dict, local_pred_dict, deep_pred_dicts, global_gt_dict_train, local_gt_dict_train, global_gt_dict_val, local_gt_dict_val, training_flag, options)
#loss_rgbd, loss_dict_rgbd, _ = build_loss_rgbd(global_pred_dict, deep_pred_dicts, global_gt_dict_rgbd_train, global_gt_dict_rgbd_val, training_flag, options)
loss, loss_dict, debug_dict = build_loss(img_inp_train, img_inp_val, global_pred_dict, deep_pred_dicts, global_gt_dict_train, global_gt_dict_val, training_flag, options)
#loss = tf.cond(tf.less(training_flag, 2), lambda: loss, lambda: tf.cond(tf.less(training_flag, 4), lambda: loss_rgbd, lambda: loss_3d))
#train_writer = tf.summary.FileWriter(options.log_dir + '/train')
#val_writer = tf.summary.FileWriter(options.log_dir + '/val')
#train_writer_rgbd = tf.summary.FileWriter(options.log_dir + '/train_rgbd')
#val_writer_rgbd = tf.summary.FileWriter(options.log_dir + '/val_rgbd')
#writers = [train_writer, val_writer, train_writer_rgbd, val_writer_rgbd]
with tf.variable_scope('statistics'):
batchno = tf.Variable(0, dtype=tf.int32, trainable=False, name='batchno')
batchnoinc=batchno.assign(batchno+1)
pass
optimizer = tf.train.AdamOptimizer(options.LR)
if options.crfrnn >= 0:
train_op = optimizer.minimize(loss, global_step=batchno)
else:
var_to_train = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, "crfrnn")
print(var_to_train)
train_op = optimizer.minimize(loss, global_step=batchno, var_list=var_to_train)
pass
var_to_restore = [v for v in tf.global_variables()]
tf.summary.scalar('loss', loss)
summary_op = tf.summary.merge_all()
config=tf.ConfigProto()
config.allow_soft_placement=True
#config.gpu_options.allow_growth=True
config.gpu_options.per_process_gpu_memory_fraction=0.9
saver=tf.train.Saver()
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
with tf.Session(config=config) as sess:
sess.run(init_op)
if options.restore == 0:
#fine-tune from DeepLab model
var_to_restore = [v for v in var_to_restore if 'res5d' not in v.name and 'segmentation' not in v.name and 'plane' not in v.name and 'deep_supervision' not in v.name and 'local' not in v.name and 'boundary' not in v.name and 'degridding' not in v.name and 'res2a_branch2a' not in v.name and 'res2a_branch1' not in v.name and 'Adam' not in v.name and 'beta' not in v.name and 'statistics' not in v.name and 'semantics' not in v.name]
pretrained_model_loader = tf.train.Saver(var_to_restore)
pretrained_model_loader.restore(sess, options.modelPathDeepLab)
elif options.restore == 1:
#restore the same model from checkpoint
loader = tf.train.Saver(var_to_restore)
loader.restore(sess,"%s/checkpoint.ckpt"%(options.checkpoint_dir))
bno=sess.run(batchno)
print(bno)
pass
# Start input enqueue threads.
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
MOVING_AVERAGE_DECAY = 0.99
ema = [0., 0., 0., 0.]
ema_acc = [1e-10, 1e-10, 1e-10, 1e-10]
last_snapshot_time = time.time()
bno=sess.run(batchno)
try:
while bno<300000:
t0 = time.time()
batchIndexPeriod = bno % 100
if batchIndexPeriod < len(options.hybrid):
#batchType = int(options.hybrid[batchIndexPeriod]) * 2 + 1
batchType = 1
_, total_loss, losses, summary_str, pred = sess.run([batchnoinc, loss, loss_dict, summary_op, global_pred_dict], feed_dict = {training_flag: batchType == 0})
else:
batchType = 0
_, total_loss, losses, summary_str, pred, debug, img, gt = sess.run([train_op, loss, loss_dict, summary_op, global_pred_dict, debug_dict, img_inp_train, global_gt_dict_train], feed_dict = {training_flag: batchType == 0})
if bno % (100 + 400 * int(options.crfrnn == 0)) == 50:
for batchIndex in xrange(options.batchSize):
#print(losses)
#print(debug['plane'][batchIndex])
cv2.imwrite('test/' + str(batchIndex) + '_image.png', ((img[batchIndex] + 0.5) * 255).astype(np.uint8))
cv2.imwrite('test/' + str(batchIndex) + '_segmentation_pred.png', drawSegmentationImage(np.concatenate([pred['segmentation'][batchIndex], pred['non_plane_mask'][batchIndex]], axis=2), blackIndex=options.numOutputPlanes))
cv2.imwrite('test/' + str(batchIndex) + '_segmentation_gt.png', drawSegmentationImage(np.concatenate([gt['segmentation'][batchIndex], gt['non_plane_mask'][batchIndex]], axis=2), blackIndex=options.numOutputPlanes))
cv2.imwrite('test/' + str(batchIndex) + '_depth.png', drawDepthImage(debug['depth'][batchIndex].squeeze()))
continue
#exit(1)
pass
pass
ema[batchType] = ema[batchType] * MOVING_AVERAGE_DECAY + total_loss
ema_acc[batchType] = ema_acc[batchType] * MOVING_AVERAGE_DECAY + 1
bno = sess.run(batchno)
if time.time()-last_snapshot_time > options.saveInterval:
print('save snapshot')
saver.save(sess,'%s/checkpoint.ckpt'%options.checkpoint_dir)
last_snapshot_time = time.time()
pass
print bno,'train', ema[0] / ema_acc[0], 'val', ema[1] / ema_acc[1], 'train rgbd', ema[2] / ema_acc[2], 'val rgbd', ema[3] / ema_acc[3], 'loss', total_loss, 'time', time.time()-t0
if np.random.random() < 0.01:
print(losses)
pass
continue
except tf.errors.OutOfRangeError:
print('Done training -- epoch limit reached')
finally:
# When done, ask the threads to stop.
coord.request_stop()
pass
# Wait for threads to finish.
coord.join(threads)
sess.close()
pass
return
def test(options):
if not os.path.exists(options.test_dir):
os.system("mkdir -p %s"%options.test_dir)
pass
if options.dataset == '':
assert(len(options.hybrid) == 1)
if options.hybrid == '0':
options.dataset = 'SUNCG'
elif options.hybrid == '1':
options.dataset = 'NYU_RGBD'
elif options.hybrid == '2':
options.dataset = 'matterport'
elif options.hybrid == '3':
options.dataset = 'ScanNet'
options.dataset
options.batchSize = 1
min_after_dequeue = 1000
reader = RecordReaderAll()
if options.dataset == 'SUNCG':
filename_queue = tf.train.string_input_producer([options.dataFolder + '/planes_SUNCG_val.tfrecords'], num_epochs=10000)
elif options.dataset == 'NYU_RGBD':
filename_queue = tf.train.string_input_producer([options.dataFolder + '/planes_nyu_rgbd_val.tfrecords'], num_epochs=1)
options.deepSupervision = 0
options.predictLocal = 0
elif options.dataset == 'matterport':
filename_queue = tf.train.string_input_producer([options.dataFolder + '/planes_matterport_val.tfrecords'], num_epochs=1)
else:
filename_queue = tf.train.string_input_producer([options.dataFolder + '/planes_scannet_val.tfrecords'], num_epochs=1)
pass
img_inp, global_gt_dict, local_gt_dict = reader.getBatch(filename_queue, numOutputPlanes=options.numOutputPlanes, batchSize=options.batchSize, min_after_dequeue=min_after_dequeue, getLocal=True, random=False)
training_flag = tf.constant(False, tf.bool)
global_pred_dict, local_pred_dict, deep_pred_dicts = build_graph(img_inp, img_inp, training_flag, options)
var_to_restore = tf.global_variables()
loss, loss_dict, debug_dict = build_loss(img_inp, img_inp, global_pred_dict, deep_pred_dicts, global_gt_dict, global_gt_dict, training_flag, options)
config=tf.ConfigProto()
config.gpu_options.allow_growth=True
config.allow_soft_placement=True
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
with tf.Session(config=config) as sess:
sess.run(init_op)
#var_to_restore = [v for v in var_to_restore if 'res4b22_relu_non_plane' not in v.name]
loader = tf.train.Saver(var_to_restore)
loader.restore(sess, "%s/checkpoint.ckpt"%(options.checkpoint_dir))
#loader.restore(sess, "%s/checkpoint.ckpt"%('checkpoint/planenet_pb_pp_hybrid1'))
#loader.restore(sess, options.fineTuningCheckpoint)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
try:
gtDepths = []
predDepths = []
planeMasks = []
#predMasks = []
gtPlanes = []
predPlanes = []
gtSegmentations = []
predSegmentations = []
gtNumPlanes = []
imageWidth = WIDTH
imageHeight = HEIGHT
focalLength = 517.97
urange = np.arange(imageWidth).reshape(1, -1).repeat(imageHeight, 0) - imageWidth * 0.5
vrange = np.arange(imageHeight).reshape(-1, 1).repeat(imageWidth, 1) - imageHeight * 0.5
ranges = np.array([urange / imageWidth * 640 / focalLength, np.ones(urange.shape), -vrange / imageHeight * 480 / focalLength]).transpose([1, 2, 0])
for index in xrange(10):
print(('image', index))
t0=time.time()
img, global_gt, local_gt, global_pred, local_pred, deep_preds, total_loss, losses, debug = sess.run([img_inp, global_gt_dict, local_gt_dict, global_pred_dict, local_pred_dict, deep_pred_dicts, loss, loss_dict, debug_dict])
# print(options.test_dir)
# cv2.imwrite(options.test_dir + '/depth.png', drawDepthImage(debug['depth'][0]))
# cv2.imwrite(options.test_dir + '/normal.png', drawNormalImage(debug['normal'][0]))
# boundary = debug['boundary'][0]
# boundary = np.concatenate([boundary, np.zeros((HEIGHT, WIDTH, 1))], axis=2)
# cv2.imwrite(options.test_dir + '/boundary.png', drawMaskImage(boundary))
# cv2.imwrite(options.test_dir + '/depth_gt.png', drawDepthImage(debug['depth_gt'][0].squeeze()))
# exit(1)
if 'pixelwise' in options.suffix:
image = ((img[0] + 0.5) * 255).astype(np.uint8)
gt_d = global_gt['depth'].squeeze()
pred_d = global_pred['non_plane_depth'].squeeze()
#depth = global_gt['depth'].squeeze()
if '_2' in options.suffix:
pred_p, pred_s, pred_d = fitPlanes(pred_d, numPlanes=options.numOutputPlanes, planeAreaThreshold=3*4, numIterations=100, distanceThreshold=0.05, local=0.2)
elif '_3' in options.suffix:
pred_p, pred_s, pred_d = fitPlanes(gt_d, numPlanes=options.numOutputPlanes, planeAreaThreshold=3*4, numIterations=100, distanceThreshold=0.05, local=0.2)
pass
# gt_p = global_gt['plane'][0]
# pred_p = planes
# valid_mask = (np.linalg.norm(gt_p, axis=1) > 0).astype(np.float32)
# diff = np.min(np.linalg.norm(np.expand_dims(gt_p, 1) - np.expand_dims(pred_p, 0), axis=2), 1)
# num += valid_mask.sum()
# lossSum += (diff * valid_mask).sum()
if options.dataset == 'SUNCG':
planeMask = np.squeeze(debug['segmentation']).sum(axis=2)
else:
planeMask = np.ones((HEIGHT, WIDTH))
if '_2' in options.suffix or '_3' in options.suffix:
planeMask *= (pred_s < options.numOutputPlanes).astype(np.float32)
pass
if index < 10:
cv2.imwrite(options.test_dir + '/' + str(index) + '_image.png', image)
cv2.imwrite(options.test_dir + '/' + str(index) + '_depth.png', drawDepthImage(gt_d))
cv2.imwrite(options.test_dir + '/' + str(index) + '_depth_pred.png', drawDepthImage(pred_d))
if '_2' in options.suffix or '_3' in options.suffix:
cv2.imwrite(options.test_dir + '/' + str(index) + '_segmentation_pred.png', drawSegmentationImage(pred_s))
pass
#cv2.imwrite(options.test_dir + '/' + str(index) + '_plane_mask.png', drawMaskImage(planeMask))
pass
gtDepths.append(gt_d)
predDepths.append(pred_d)
planeMasks.append(planeMask)
if options.dataset != 'NYU_RGBD' and ('_2' in options.suffix or '_3' in options.suffix):
gt_p = global_gt['plane'][0]
gt_s = global_gt['segmentation'][0]
gt_num_p = global_gt['num_planes'][0]
pred_s = (np.expand_dims(pred_s, -1) == np.reshape(np.arange(options.numOutputPlanes), [1, 1, -1])).astype(np.float32)
gtPlanes.append(gt_p)
predPlanes.append(pred_p)
gtSegmentations.append(gt_s)
gtNumPlanes.append(gt_num_p)
predSegmentations.append(pred_s)
pass
#planeMasks.append((planeSegmentation < options.numOutputPlanes).astype(np.float32))
continue
print(losses)
print(total_loss)
#print(losses)
#exit(1)
im = img[0]
image = ((im + 0.5) * 255).astype(np.uint8)
gt_d = global_gt['depth'].squeeze()
if options.predictLocal:
pred_local_s = 1 / (1 + np.exp(-local_pred['score'][0]))
pred_local_p = local_pred['plane'][0]
pred_local_m = local_pred['mask'][0]
gt_local_s = local_gt['score'][0]
gt_local_m = local_gt['mask'][0]
#visualize local plane prediction
stride = 8
boxSize = 64
xs = np.arange(WIDTH / stride) * stride + stride / 2
ys = np.arange(HEIGHT / stride) * stride + stride / 2
padding = boxSize / 2 + 1
maskImage = np.zeros((HEIGHT + padding * 2, WIDTH + padding * 2, 3), dtype=np.uint8)
maskImage[padding:padding + HEIGHT, padding:padding + WIDTH, :] = image / 2
for gridY, y in enumerate(ys):
for gridX, x in enumerate(xs):
score = pred_local_s[gridY][gridX]
if score < 0.5:
continue
mask = pred_local_m[gridY][gridX].reshape([16, 16])
mask = cv2.resize(mask, (boxSize, boxSize))
maskImage[y - boxSize / 2 + padding:y + boxSize / 2 + padding, x - boxSize / 2 + padding:x + boxSize / 2 + padding, 0][mask > 0.5] = 255
continue
continue
for gridY, y in enumerate(ys):
for gridX, x in enumerate(xs):
score = gt_local_s[gridY][gridX]
if score < 0.5:
continue
mask = gt_local_m[gridY][gridX].reshape([16, 16])
mask = cv2.resize(mask, (boxSize, boxSize))
maskImage[y - boxSize / 2 + padding:y + boxSize / 2 + padding, x - boxSize / 2 + padding:x + boxSize / 2 + padding, 2][mask > 0.5] = 255
continue
continue
pass
pred_p = global_pred['plane'][0]
pred_s = global_pred['segmentation'][0]
pred_np_m = global_pred['non_plane_mask'][0]
pred_np_d = global_pred['non_plane_depth'][0]
pred_np_n = global_pred['non_plane_normal'][0]
planeMask = 1 - global_gt['non_plane_mask'][0]
info = global_gt['info'][0]
all_segmentations = np.concatenate([pred_s, pred_np_m], axis=2)
all_segmentations_softmax = softmax(all_segmentations)
segmentation = np.argmax(all_segmentations, 2)
#pred_p, segmentation, numPlanes = mergePlanes(global_gt['plane'][0], np.concatenate([global_gt['segmentation'][0], global_gt['non_plane_mask'][0]], axis=2), global_gt['depth'][0].squeeze(), global_gt['info'][0], np.concatenate([pred_s, pred_np_m], axis=2))
plane_depths = calcPlaneDepths(pred_p, WIDTH, HEIGHT, info)
all_depths = np.concatenate([plane_depths, pred_np_d], axis=2)
pred_d = all_depths.reshape(-1, options.numOutputPlanes + 1)[np.arange(WIDTH * HEIGHT), segmentation.reshape(-1)].reshape(HEIGHT, WIDTH)
plane_normals = calcPlaneNormals(pred_p, WIDTH, HEIGHT)
all_normals = np.concatenate([plane_normals, np.expand_dims(pred_np_n, 2)], axis=2)
pred_n = np.sum(all_normals * np.expand_dims(one_hot(segmentation, options.numOutputPlanes+1), -1), 2)
#pred_n = all_normals.reshape(-1, options.numOutputPlanes + 1, 3)[np.arange(WIDTH * HEIGHT), segmentation.reshape(-1)].reshape((HEIGHT, WIDTH, 3))
if False:
gt_s = global_gt['segmentation'][0]
all_segmentations = np.concatenate([gt_s, 1 - planeMask], axis=2)
gt_p = global_gt['plane'][0]
# valid_mask = (np.linalg.norm(gt_p, axis=1) > 0).astype(np.float32)
# diff = np.min(np.linalg.norm(np.expand_dims(gt_p, 1) - np.expand_dims(pred_p, 0), axis=2), 1)
# num += valid_mask.sum()
# lossSum += (diff * valid_mask).sum()
#gt_p = np.stack([-gt_p[:, 0], -gt_p[:, 2], -gt_p[:, 1]], axis=1)
plane_depths = calcPlaneDepths(gt_p, WIDTH, HEIGHT, info)
all_depths = np.concatenate([plane_depths, pred_np_d], axis=2)
segmentation = np.argmax(all_segmentations, 2)
pred_d = all_depths.reshape(-1, options.numOutputPlanes + 1)[np.arange(WIDTH * HEIGHT), segmentation.reshape(-1)].reshape(HEIGHT, WIDTH)
# print(gt_p)
# for segmentIndex in xrange(options.numOutputPlanes):
# cv2.imwrite(options.test_dir + '/' + str(index) + '_mask_' + str(segmentIndex) + '.png', drawMaskImage(segmentation == segmentIndex))
# continue
cv2.imwrite(options.test_dir + '/' + str(index) + '_depth_pred.png', drawDepthImage(pred_d))
#exit(1)
pass
gt_p = global_gt['plane'][0]
gt_s = global_gt['segmentation'][0]
gt_num_p = global_gt['num_planes'][0]
gtPlanes.append(gt_p)
predPlanes.append(pred_p)
gtSegmentations.append(gt_s)
gtNumPlanes.append(gt_num_p)
predSegmentations.append(pred_s)
gtDepths.append(gt_d)
planeMasks.append(planeMask.squeeze())
predDepths.append(pred_d)
evaluateDepths(predDepths[-1], gtDepths[-1], gtDepths[-1] > 0, planeMasks[-1])
if index >= 10:
continue
if options.predictSemantics:
cv2.imwrite(options.test_dir + '/' + str(index) + '_semantics_pred.png', drawSegmentationImage(global_pred['semantics'][0], blackIndex=0))
cv2.imwrite(options.test_dir + '/' + str(index) + '_semantics_gt.png', drawSegmentationImage(global_gt['semantics'][0], blackIndex=0))
pass
if 'cost_mask' in debug:
cv2.imwrite(options.test_dir + '/' + str(index) + '_cost_mask.png', drawMaskImage(np.sum(debug['cost_mask'][0], axis=-1)))
for planeIndex in xrange(options.numOutputPlanes + 1):
cv2.imwrite(options.test_dir + '/' + str(index) + '_cost_mask_' + str(planeIndex) + '.png', drawMaskImage(debug['cost_mask'][0, :, :, planeIndex]))
continue
all_segmentations = np.concatenate([pred_s, pred_np_m], axis=2)
for planeIndex in xrange(options.numOutputPlanes + 1):
cv2.imwrite(options.test_dir + '/' + str(index) + '_segmentation_pred_' + str(planeIndex) + '.png', drawMaskImage(all_segmentations[:, :, planeIndex]))
continue
exit(1)
pass
if 'normal' in global_gt:
gt_n = global_gt['normal'][0]
norm = np.linalg.norm(gt_n, axis=-1, keepdims=True)
gt_n /= np.maximum(norm, 1e-4)
#gt_n = np.stack([-gt_n[:, :, 0], -gt_n[:, :, 2], -gt_n[:, :, 1]], axis=2)
cv2.imwrite(options.test_dir + '/' + str(index) + '_normal_gt.png', drawNormalImage(gt_n))
#cv2.imwrite(options.test_dir + '/' + str(index) + '_normal_pred.png', drawNormalImage(pred_np_n))
cv2.imwrite(options.test_dir + '/' + str(index) + '_normal_pred.png', drawNormalImage(pred_n))
pass
if 'segmentation' in global_gt:
gt_s = global_gt['segmentation'][0]
gt_p = global_gt['plane'][0]
#for planeIndex in xrange(options.numOutputPlanes):
#cv2.imwrite('test/mask_' + str(planeIndex) + '.png', drawMaskImage(gt_s[:, :, planeIndex]))
#continue
# print(gt_p)
# print(gt_n[109][129])
# print(gt_n[166][245])
# print(gt_s[109][129])
# print(gt_s[166][245])
# print(plane_normals[109][129])
# print(plane_normals[166][245])
# for planeIndex in xrange(options.numOutputPlanes):
# cv2.imwrite('test/mask_' + str(planeIndex) + '.png', drawMaskImage(gt_s[:, :, planeIndex]))
# continue
# exit(1)
gt_s, gt_p = sortSegmentations(gt_s, gt_p, pred_p)
cv2.imwrite(options.test_dir + '/' + str(index) + '_plane_mask_gt.png', drawMaskImage(planeMask))
cv2.imwrite(options.test_dir + '/' + str(index) + '_segmentation_gt.png', drawSegmentationImage(np.concatenate([gt_s, 1 - planeMask], axis=2), blackIndex=options.numOutputPlanes)
)
#cv2.imwrite(options.test_dir + '/' + str(index) + '_test.png', drawMaskImage(np.sum(np.concatenate([gt_s, 1 - planeMask], axis=2), axis=2)))
#exit(1)
#exit(1)
pass
if options.predictConfidence == 1 and options.dataset == 'SUNCG':
assert(False)
pred_p_c = global_pred['confidence'][0]
pred_p_c = 1 / (1 + np.exp(-pred_p_c))
#print(pred_p_c)
# print(losses)
# print(debug['plane'][0])
# print(pred_p)
# exit(1)
numPlanes = global_gt['num_planes'][0]
print((numPlanes, (pred_p_c > 0.5).sum()))
pred_p_c = (pred_p_c > 0.5).astype(np.float32)
pred_p *= pred_p_c
pred_s -= (1 - pred_p_c.reshape([1, 1, options.numOutputPlanes])) * 10
pass
if False:
#dump results for post processing
if index >= 10:
break
np.save(options.dump_dir + '/planes_' + str(index) + '.npy', pred_p)
np.save(options.dump_dir + '/segmentations_' + str(index) + '.npy', pred_s)
np.save(options.dump_dir + '/segmentations_gt_' + str(index) + '.npy', gt_s)
np.save(options.dump_dir + '/non_plane_depth_' + str(index) + '.npy', pred_np_d)
np.save(options.dump_dir + '/non_plane_segmentation_' + str(index) + '.npy', pred_np_m)
boundary = pred_boundary[0]
boundary = 1 / (1 + np.exp(-boundary))
boundary = np.concatenate([boundary, np.zeros((HEIGHT, WIDTH, 1))], axis=2)
cv2.imwrite(options.dump_dir + '/boundary_' + str(index) + '.png', drawMaskImage(boundary))
cv2.imwrite(options.dump_dir + '/image_' + str(index) + '.png', cv2.resize(image, (WIDTH, HEIGHT), interpolation=cv2.INTER_LINEAR))
np.save(options.dump_dir + '/depth_' + str(index) + '.npy', gt_d)
continue
cv2.imwrite(options.test_dir + '/' + str(index) + '_image.png', image)
cv2.imwrite(options.test_dir + '/' + str(index) + '_depth.png', drawDepthImage(gt_d))
#cv2.imwrite(options.test_dir + '/' + str(index) + '_overlay.png', drawDepthImageOverlay(image, gt_d))
if options.predictBoundary:
pred_boundary = global_pred['boundary'][0]
pred_boundary = 1 / (1 + np.exp(-pred_boundary))
boundary = np.concatenate([pred_boundary, np.zeros((HEIGHT, WIDTH, 1))], axis=2)
cv2.imwrite(options.test_dir + '/' + str(index) + '_boundary_pred.png', drawMaskImage(boundary))
pass
if 'boundary' in global_gt:
gt_boundary = global_gt['boundary'][0]
boundary = np.concatenate([gt_boundary, np.zeros((HEIGHT, WIDTH, 1))], axis=2)
cv2.imwrite(options.test_dir + '/' + str(index) + '_boundary_gt.png', drawMaskImage(boundary))
pass
for layerIndex, layer in enumerate(options.deepSupervisionLayers):
segmentation_deep = np.argmax(deep_preds[layerIndex]['segmentation'][0], 2)
segmentation_deep[segmentation_deep == options.numOutputPlanes] = -1
segmentation_deep += 1
plane_depths_deep = calcPlaneDepths(deep_preds[layerIndex]['plane'][0], WIDTH, HEIGHT, info)
all_depths_deep = np.concatenate([pred_np_d, plane_depths_deep], axis=2)
pred_d_deep = all_depths_deep.reshape(-1, options.numOutputPlanes + 1)[np.arange(WIDTH * HEIGHT), segmentation_deep.reshape(-1)].reshape(HEIGHT, WIDTH)
cv2.imwrite(options.test_dir + '/' + str(index) + '_segmentation_pred_' + str(layerIndex) + '.png', drawSegmentationImage(deep_preds[layerIndex]['segmentation'][0]))
cv2.imwrite(options.test_dir + '/' + str(index) + '_depth_pred_' + str(layerIndex) + '.png', drawDepthImage(pred_d_deep))
pass
#print(pred_np_m)
#print(pred_s)