-
Notifications
You must be signed in to change notification settings - Fork 11
/
dataloader.py
1063 lines (741 loc) · 38.8 KB
/
dataloader.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 math
import cv2
import torch
from torch.utils.data.dataset import Dataset
# General
from PIL import Image
import numpy as np
import os.path
import scipy.io
import copy
import pickle
import struct
from scipy import stats, ndimage
DATASET_LENGTHS={};
DATASET_LENGTHS["nyu"] = 72757
DATASET_LENGTHS["icvl"] = 22067
DATASET_LENGTHS["msra"] = 76375 - 8500
DATASET_NUM_JOINTS = {};
DATASET_NUM_JOINTS["nyu"]=14
DATASET_NUM_JOINTS["icvl"]=16
DATASET_NUM_JOINTS["msra"]=21
class HandPoseDataset(Dataset):
def __init__(self,
basepath="",train=True,cropSize=(128,128),doJitterRotation=False,doAddWhiteNoise=False,rotationAngleRange=[-45.0, 45.0],
comJitter=False,RandDotPercentage=0, indeces=None, sigmaNoise=2,cropSize3D=[250,250,250],
do_norm_zero_one=False,random_seed=21,drop_joint_num=0,center_refined=False,horizontal_flip=0, scale_aug=0):
# do_norm_zero_one if is False, the depth values will be squashed to the interval [-1,+1]
self.min_depth_cam = 50.
self.max_depth_cam = 1500.
self.do_norm_zero_one=do_norm_zero_one
self.doJitterRotation=doJitterRotation
self.rotationAngleRange=rotationAngleRange
self.basepath = basepath
self.restrictedJointsEval = nyuRestrictedJointsEval
self.cropSize3D = cropSize3D
self.RandDotPercentage = RandDotPercentage
self.drop_joint_num=drop_joint_num # num joints to be dropped for each frame
self.horizontal_flip=horizontal_flip # random horizontal flip
# For comparisons check results with adapted cube size
self.testseq2_start_id = 2441
self.cropSize = cropSize
self.doAddWhiteNoise = doAddWhiteNoise
self.sigmaNoise = sigmaNoise
self.doNormZeroOne = do_norm_zero_one # [-1,1] or [0,1]
self.center_refined=center_refined
self.scale_aug=scale_aug
self.randomAngle=0
self.randomScale=1
self.randomComJitter=0 * np.clip(np.random.randn(3)*6,-6,6)
self.comJitter = comJitter
self.doTrain = train
self.seqName = ""
if self.doTrain:
self.seqName = "train"
else:
self.seqName = "test"
if indeces is None:
self.numSamples=self.numSamples
self.indeces = [i for i in range(self.numSamples)]
else:
self.indeces = indeces
self.numSamples = len(indeces)
if drop_joint_num>0:
print(f"{drop_joint_num} out of {self.num_joints} joints will be dropped for each frame")
joint_list=[i for i in range(self.num_joints)]
np.random.seed(random_seed+10)
self.dropped_joints=[np.random.choice(joint_list,drop_joint_num,replace=False) for j in range(self.numSamples)]
def __len__(self):
return self.numSamples
def get_validIndex(self,ind):
"""
This function should be overwritten in subclasses
"""
raise NotImplementedError()
def LoadSample(self,ind,com=None):
"""
This function should be overwritten in subclasses
"""
raise NotImplementedError()
def convert_uvd_to_xyz_tensor(self,uvd):
"""
This function should be overwritten in subclasses
"""
raise NotImplementedError()
def convert_xyz_to_uvd_tensor(self,xyz):
"""
This function should be overwritten in subclasses
"""
raise NotImplementedError()
def pointImgTo3D(self,sample):
"""
This function should be overwritten in subclasses
"""
raise NotImplementedError()
def __getitem__(self, index):
valIndex = self.get_validIndex(index)
self.valIndex=valIndex
if self.center_refined:
com=self.center_refined_uvd[valIndex]
else:
com=None
data = self.LoadSample(valIndex,com)
if self.doNormZeroOne:
img, target = normalizeZeroOne(data)
else:
img, target = normalizeMinusOneOne(data)
self.sample_loaded = data
com = torch.from_numpy(data["com3D"])
gt2Dcrop = torch.from_numpy((data["gt2Dcrop"]-np.array([0,0,com[2]]))/np.array([1,1.,data['cubesize'][2] / 2.]))
if self.horizontal_flip and np.random.rand()<0.5: # remember in this case, M will not result in the original UVD
img,gt2Dcrop=horizontal_flip_depth(img,gt2Dcrop)
if self.scale_aug and np.random.rand()<0.7: # remember in this case, M will not result in the original UVD
scale = 0.8 + np.random.random() * 0.4
img,gt2Dcrop=scale_depth(img,gt2Dcrop,scale)
self.randomScale=scale
else:
self.randomScale=1
# Image need to be HxWxC and will be divided by transform (ToTensor()), which is assumed here!
img = np.expand_dims(img, axis=0)
img=torch.from_numpy(img)
if self.RandDotPercentage>0 and np.random.rand()<0.7:
p=np.random.rand()*self.RandDotPercentage
img=PixDropout(img,background_value=1,P=p,V=1)
#target = torch.from_numpy(target.astype('float32'))self. = RandDotPercentage
M_=torch.from_numpy(data["M"])
M_=torch.cat( [ torch.cat([ M_[:,:2],torch.zeros(3,1),M_[:,2][...,None]],dim=-1), torch.zeros(1,4)]);M_[2,2]=1;M_[2,3]=0;
M=M_.float()
M_inv=torch.from_numpy(np.linalg.inv(data["M"]))
M_=torch.cat( [ torch.cat([ M_inv[:,:2],torch.zeros(3,1),M_inv[:,2][...,None]],dim=-1), torch.zeros(1,4)]);M_[2,2]=1;M_[2,3]=0;
M_inv=M_.float()
cubesize = torch.from_numpy(np.array(data["cubesize"])).float()
gt2Dorignal = torch.from_numpy(data["gt2Dorignal"]).float()
gt3Dorignal = torch.from_numpy(data["gt3Dorignal"]).float()
visible_mask=get_visible(img , gt2Dcrop[:,:2], cropSize=self.cropSize[0], background_value=1) # joints that are visible within the image frame
#get_visible(gt2Dcrop[:,:2],cropSize=self.cropSize[0])
joint_mask=torch.ones(self.num_joints,1, dtype=torch.bool)
if self.drop_joint_num>0:
joint_mask[self.dropped_joints[index]]=False
return img.float(), gt2Dcrop.float(), gt2Dorignal, gt3Dorignal, com.float(), M_inv, cubesize.float(), joint_mask.float(), visible_mask, M
def cropArea3D(self,imgDepth, com, minRatioInside=0.75, size=(250, 250, 250), dsize=(128, 128)):
"""
Crop area of hand in 3D volumina, scales inverse to the distance of hand to camera
:param com: center of mass, in image coordinates (x,y,z), z in mm
:param size: (x,y,z) extent of the source crop volume in mm
:param dsize: (x,y) extent of the destination size
:return: cropped hand image, transformation matrix for joints, CoM in image coordinates
"""
RESIZE_BILINEAR = 0
RESIZE_CV2_NN = 1
RESIZE_CV2_LINEAR = 2
CROP_BG_VALUE = 0.0
resizeMethod = RESIZE_CV2_NN
# calculate boundaries
xstart, xend, ystart, yend, zstart, zend = comToBounds(com.copy(), size, self.fx , self.fy)
# Check if part within image is large enough; otherwise stop
xstartin = max(xstart,0)
xendin = min(xend, imgDepth.shape[1])
ystartin = max(ystart,0)
yendin = min(yend, imgDepth.shape[0])
ratioInside = float((xendin - xstartin) * (yendin - ystartin)) / float((xend - xstart) * (yend - ystart))
if (ratioInside < minRatioInside) \
and ((com[0] < 0) \
or (com[0] >= imgDepth.shape[1]) \
or (com[1] < 0) or (com[1] >= imgDepth.shape[0])):
print("Hand largely outside image (ratio (inside) = {})".format(ratioInside))
raise UserWarning('Hand not inside image')
# crop patch from source
cropped = imgDepth[max(ystart, 0):min(yend, imgDepth.shape[0]),
max(xstart, 0):min(xend, imgDepth.shape[1])].copy()
# add pixels that are out of the image in order to keep aspect ratio
cropped = np.pad(cropped, ((abs(ystart)-max(ystart, 0), abs(yend)-min(yend, imgDepth.shape[0])),
(abs(xstart)-max(xstart, 0), abs(xend)-min(xend, imgDepth.shape[1]))),
mode='constant', constant_values=int(CROP_BG_VALUE))
msk1 = np.bitwise_and(cropped < zstart, cropped != 0)
msk2 = np.bitwise_and(cropped > zend, cropped != 0)
# Backface is at 0, it is set later;
# setting anything outside cube to same value now (was set to zstart earlier)
cropped[msk1] = CROP_BG_VALUE
cropped[msk2] = CROP_BG_VALUE
wb = (xend - xstart)
hb = (yend - ystart)
trans = np.asmatrix(np.eye(3, dtype=float))
trans[0, 2] = -xstart
trans[1, 2] = -ystart
# Compute size of image patch for isotropic scaling
# where the larger side is the side length of the fixed size image patch (preserving aspect ratio)
if wb > hb:
sz = (dsize[0], int(round(hb * dsize[0] / float(wb))))
else:
sz = (int(round(wb * dsize[1] / float(hb))), dsize[1])
# Compute scale factor from cropped ROI in image to fixed size image patch;
# set up matrix with same scale in x and y (preserving aspect ratio)
roi = cropped
if roi.shape[0] > roi.shape[1]: # Note, roi.shape is (y,x) and sz is (x,y)
scale = np.asmatrix(np.eye(3, dtype=float) * sz[1] / float(roi.shape[0]))
else:
scale = np.asmatrix(np.eye(3, dtype=float) * sz[0] / float(roi.shape[1]))
scale[2, 2] = 1
# depth resize
if resizeMethod == RESIZE_CV2_NN:
rz = cv2.resize(cropped, sz, interpolation=cv2.INTER_NEAREST)
elif resizeMethod == RESIZE_BILINEAR:
rz = HandDetector.bilinearResize(cropped, sz, CROP_BG_VALUE)
elif resizeMethod == RESIZE_CV2_LINEAR:
rz = cv2.resize(cropped, sz, interpolation=cv2.INTER_LINEAR)
else:
raise NotImplementedError("Unknown resize method!")
# Sanity check
# numValidPixels = np.sum(rz != CROP_BG_VALUE)
# if (numValidPixels < 40) or (numValidPixels < (np.prod(dsize) * 0.01)):
# print("Too small number of foreground/hand pixels: {}/{} ({}))".format(
# numValidPixels, np.prod(dsize), dsize))
# raise UserWarning("No valid hand. Foreground region too small.")
# Place the resized patch (with preserved aspect ratio)
# in the center of a fixed size patch (padded with default background values)
ret = np.ones(dsize, np.float32) * CROP_BG_VALUE # use background as filler
xstart = int(math.floor(dsize[0] / 2 - rz.shape[1] / 2))
xend = int(xstart + rz.shape[1])
ystart = int(math.floor(dsize[1] / 2 - rz.shape[0] / 2))
yend = int(ystart + rz.shape[0])
ret[ystart:yend, xstart:xend] = rz
off = np.asmatrix(np.eye(3, dtype=float))
off[0, 2] = xstart
off[1, 2] = ystart
return ret, off * scale * trans, com
nyuRestrictedJointsEval = [0, 3, 6, 9, 12, 15, 18, 21, 24, 25, 27, 30, 31, 32]
class NYUHandPoseDataset(HandPoseDataset):
def __init__(self, basepath="",train=True,cropSize=(128,128),doJitterRotation=False,doAddWhiteNoise=False,rotationAngleRange=[-45.0, 45.0],
comJitter=False,RandDotPercentage=0, indeces=None, sigmaNoise=1,cropSize3D=[250,250,250],camID=1,
do_norm_zero_one=False,doLoadRealSample=True,random_seed=21,drop_joint_num=0,center_refined=False,horizontal_flip=0, scale_aug=0):
self.fx, self.fy, self.ux , self.uy = (588.036865, 587.075073, 320, 240)
self.camID=camID;
self.doLoadRealSample=doLoadRealSample
self.num_joints = len(nyuRestrictedJointsEval)
if train:
self.seqName = "train"
else:
self.seqName = "test"
# Load labels
labels = '{}/{}/joint_data.mat'.format(basepath, self.seqName)
self.labelMat = scipy.io.loadmat(labels)
# Get number of samples from annotations (test: 8252; train: 72757)
self.numSamples = self.labelMat['joint_xyz'][camID-1].shape[0]
super(NYUHandPoseDataset, self).__init__(basepath=basepath,train=train,cropSize=cropSize,doJitterRotation=doJitterRotation,doAddWhiteNoise=doAddWhiteNoise,rotationAngleRange=rotationAngleRange,
comJitter=comJitter,RandDotPercentage=RandDotPercentage, indeces=indeces, sigmaNoise=sigmaNoise,cropSize3D=cropSize3D,
do_norm_zero_one=do_norm_zero_one,random_seed=random_seed,drop_joint_num=drop_joint_num,center_refined=center_refined,horizontal_flip=horizontal_flip, scale_aug=scale_aug)
if center_refined:
print("Center refined being used")
center_path = os.path.join(basepath,'center_{}_refined.txt'.format(self.seqName))
refined=np.loadtxt(center_path)
self.center_refined_uvd = np.array(refined)# self.convert_xyz_to_uvd_tensor(torch.tensor(refined).unsqueeze(1)) ).squeeze() #(B,1, 3) -> (B,3)
print("NYU Dataset init done.")
def LoadSample(self,ind,com=None):
idComGT = 13
# Load the dataset
objdir = '{}/{}/'.format(self.basepath,self.seqName)
if self.labelMat == None:
labelsAdress = '{}/{}/joint_data.mat'.format(self.basepath, self.seqName)
self.labelMat = scipy.io.loadmat(labelsAdress)
joints3D = self.labelMat['joint_xyz'][self.camID-1]
joints2D = self.labelMat['joint_uvd'][self.camID-1]
eval_idxs = nyuRestrictedJointsEval
numJoints = len(eval_idxs)
data = []
line = ind
# Assemble original filename
prefix = "depth" if self.doLoadRealSample else "synthdepth"
dptFileName = '{0:s}/{1:s}_{2:1d}_{3:07d}.png'.format(objdir, prefix, self.camID, line+1)
dpt = loadDepthMap(dptFileName)
# Add noise?
if self.doAddWhiteNoise:
img_white_noise_scale = np.random.randn(dpt.shape[0], dpt.shape[1])
dpt = dpt + sigmaNoise * self.img_white_noise_scale
# joints in image coordinates
gt2Dorignal = np.zeros((numJoints, 3), np.float32)
jt = 0
for ii in range(joints2D.shape[1]):
if ii not in eval_idxs:
continue
gt2Dorignal[jt,0] = joints2D[line,ii,0]
gt2Dorignal[jt,1] = joints2D[line,ii,1]
gt2Dorignal[jt,2] = joints2D[line,ii,2]
jt += 1
# normalized joints in 3D coordinates
gt3Dorignal = np.zeros((numJoints,3),np.float32)
jt = 0
for jj in range(joints3D.shape[1]):
if jj not in eval_idxs:
continue
gt3Dorignal[jt,0] = joints3D[line,jj,0]
gt3Dorignal[jt,1] = joints3D[line,jj,1]
gt3Dorignal[jt,2] = joints3D[line,jj,2]
jt += 1
#comGT = copy.deepcopy(gt2Dorignal[idComGT]) # use GT position for comparison
self.randomComJitter = np.clip(np.random.randn(3)*6,-self.comJitter,self.comJitter)#(1 if self.comJitter else 0) * np.clip(np.random.randn(3)*6,-self.comJitter,self.comJitter)
comGT = (copy.deepcopy(gt2Dorignal[idComGT]) if com is None else com) + self.randomComJitter #np.clip(np.concatenate((np.random.randn(2),np.array([0.])))*6,-6,6)
if self.doJitterRotation:
rotation_angle_scale = np.random.randn()
rot = rotation_angle_scale * (self.rotationAngleRange[1] - self.rotationAngleRange[0]) + self.rotationAngleRange[0]
self.randomAngle=rot
dpt, gt2Dorignal = rotateImageAndGt(dpt, comGT, rot, gt2Dorignal, bgValue=10000)
# Jitter scale (cube size)?
cubesize = np.float32( np.array(self.cropSize3D) * (5/6. if (not self.doTrain and ind>=2440) else 1) )
dpt, M, com = self.cropArea3D(imgDepth=dpt,com=comGT,minRatioInside=0.6 ,size=cubesize, dsize=self.cropSize)
com3D = self.pointImgTo3D(com)
gt3Dcrop = gt3Dorignal - com3D # normalize to com
gt2Dcrop = np.zeros((gt2Dorignal.shape[0], 3), np.float32)
for joint in range(gt2Dorignal.shape[0]):
t=transformPoint2D(gt2Dorignal[joint], M)
gt2Dcrop[joint, 0] = t[0]
gt2Dcrop[joint, 1] = t[1]
gt2Dcrop[joint, 2] = gt2Dorignal[joint, 2]
D={};D["M"]=M;D["com3D"]=com3D;D["cubesize"]=cubesize
D["dpt"]=dpt.astype(np.float32);D["gt2Dorignal"]=gt2Dorignal
D["gt2Dcrop"]=gt2Dcrop;D["gt3Dorignal"]=gt3Dorignal;D["gt3Dcrop"]=gt3Dcrop;
return D
def get_validIndex(self,ind):
return self.indeces[ind]
def convert_uvd_to_xyz_tensor(self, uvd ):
# uvd is a tensor of size(B,num_joints,3)
xRes = 640;
yRes = 480;
xzFactor = 1.08836710; #xzFactor=640/coeffX
yzFactor = 0.817612648;
normalizedX = uvd[:,:,0] / xRes - 0.5;
normalizedY = 0.5 - uvd[:,:,1] / yRes;
xyz = torch.zeros(uvd.shape);
xyz[:,:,2] = uvd[:,:,2];
xyz[:,:,0] = normalizedX * xyz[:,:,2] * xzFactor;
xyz[:,:,1] = normalizedY * xyz[:,:,2] * yzFactor;
return xyz
def convert_xyz_to_uvd_tensor(self, xyz):
uvd = torch.zeros(xyz.shape);
uvd[:,:,2] = xyz[:,:,2];
uvd[:,:,0] = xyz[:,:,0]/xyz[:,:,2]*self.fx+self.ux
uvd[:,:,1] = self.uy-xyz[:,:,1]/xyz[:,:,2]*self.fy
return uvd
def pointImgTo3D(self, sample):
ret = np.zeros((3,), np.float32)
# convert to metric using f, see Thomson et al.
ret[0] = (sample[0] - self.ux) * sample[2] / self.fx
ret[1] = (self.uy - sample[1]) * sample[2] / self.fy
ret[2] = sample[2]
return ret
########## ICVL ##########################################
InvalidIndicies=[6635, 8001, 9990, 10292, 12378, 19770, 19863, 20531] + [7037, 8724, 11852, 19161, 21080 , 10463 , 12837 , 19864 , 20343 , 20532 ] # the second for the case where we have center_jitter
class ICVLHandPoseDataset(HandPoseDataset):
def __init__(self, basepath="",train=True,cropSize=(128,128),doJitterRotation=False,doAddWhiteNoise=False,rotationAngleRange=[-45.0, 45.0],
comJitter=False,RandDotPercentage=0, indeces=None, sigmaNoise=1,cropSize3D=[250,250,250], do_norm_zero_one=False,random_seed=21,
drop_joint_num=0,center_refined=False,horizontal_flip=0, scale_aug=0):
self.fx, self.fy, self.ux , self.uy = (240.99, 240.96, 160, 120)#(241.42, 241.42, 160., 120.)
self.num_joints = 16
self.seqName = ""
if train:
self.seqName = "train"
else:
self.seqName = "test"
address=os.path.join(basepath,"%s.pickle"%(self.seqName))
data=pickle.load(open(address, "rb"))
self.data=data[0]
self.numSamples = len(self.data)
super(ICVLHandPoseDataset, self).__init__(basepath=basepath,train=train,cropSize=cropSize,doJitterRotation=doJitterRotation,doAddWhiteNoise=doAddWhiteNoise,rotationAngleRange=rotationAngleRange,
comJitter=comJitter,RandDotPercentage=RandDotPercentage, indeces=indeces, sigmaNoise=sigmaNoise,cropSize3D=cropSize3D,
do_norm_zero_one=do_norm_zero_one,random_seed=random_seed,drop_joint_num=drop_joint_num,center_refined=center_refined,horizontal_flip=horizontal_flip, scale_aug=scale_aug)
if center_refined:
print("Center refined being used")
self.center_refined_uvd=data[1]
print("ICVL Dataset init done.")
def get_validIndex(self,ind):
valIndex = self.indeces[ind]
while valIndex in InvalidIndicies:
valIndex=self.indeces[ np.random.randint(0,self.numSamples) ]
return valIndex
def LoadSample(self,ind,com=None):
idComGT = 0
# Load the dataset
sample=self.data[ind]
gt3Dorignal = sample[2]
gt2Dorignal = sample[1]
numJoints = gt2Dorignal.shape[0]
dpt = sample[0]
self.randomComJitter = np.clip(np.random.randn(3)*6,-self.comJitter,self.comJitter)#(1 if self.comJitter else 0)* np.clip(np.random.randn(3)*6,-6,6)
comGT = (copy.deepcopy(gt2Dorignal[idComGT]) if com is None else com) + self.randomComJitter #np.clip(np.concatenate((np.random.randn(2),np.array([0.])))*6,-6,6)
# Add noise?
if self.doAddWhiteNoise:
img_white_noise_scale = np.random.randn(dpt.shape[0], dpt.shape[1])
dpt = dpt + sigmaNoise * self.img_white_noise_scale
if self.doJitterRotation:
rotation_angle_scale = np.random.randn()
rot = rotation_angle_scale * (self.rotationAngleRange[1] - self.rotationAngleRange[0]) + self.rotationAngleRange[0]
self.randomAngle=rot
dpt, gt2Dorignal = rotateImageAndGt(dpt, comGT, rot, gt2Dorignal, bgValue=10000)
# Jitter scale (cube size)?
cubesize = self.cropSize3D
dpt, M, com = self.cropArea3D(imgDepth=dpt,com=comGT,minRatioInside=0.6,size=cubesize, dsize=self.cropSize)
com3D = self.pointImgTo3D(com)
gt3Dcrop = gt3Dorignal - com3D # normalize to com
gt2Dcrop = np.zeros((gt2Dorignal.shape[0], 3), np.float32)
for joint in range(gt2Dorignal.shape[0]):
t=transformPoint2D(gt2Dorignal[joint], M)
gt2Dcrop[joint, 0] = t[0]
gt2Dcrop[joint, 1] = t[1]
gt2Dcrop[joint, 2] = gt2Dorignal[joint, 2]
D={};D["M"]=M;D["com3D"]=com3D;D["cubesize"]=cubesize
D["dpt"]=dpt.astype(np.float32);D["gt2Dorignal"]=gt2Dorignal;D["filename"]=sample[-1]
D["gt2Dcrop"]=gt2Dcrop;D["gt3Dorignal"]=gt3Dorignal;D["gt3Dcrop"]=gt3Dcrop;
return D
def convert_uvd_to_xyz_tensor( self, uvd ):
# uvd is a tensor of size(B,num_joints,3)
xyz = torch.zeros(uvd.shape);
xyz[:,:,2] = uvd[:,:,2];
xyz[:,:,0] = (uvd[:,:,0]-self.ux)*uvd[:,:,2]/self.fx
xyz[:,:,1] = (uvd[:,:,1]-self.uy)*uvd[:,:,2]/self.fy
return xyz
def convert_xyz_to_uvd_tensor(self, xyz):
# xyz is a tensor of size(B,num_joints,3)
uvd = torch.zeros(xyz.shape);
uvd[:,:,2] = xyz[:,:,2];
uvd[:,:,0] = xyz[:,:,0]/xyz[:,:,2]*self.fx+self.ux
uvd[:,:,1] = xyz[:,:,1]/xyz[:,:,2]*self.fy+self.uy
return uvd
def pointImgTo3D(self, sample):
"""
Normalize sample to metric 3D
:param sample: joints in (x,y,z) with x,y in image coordinates and z in mm
:return: normalized joints in mm
"""
ret = np.zeros((3,), np.float32)
# convert to metric using f
ret[0] = (sample[0]-self.ux)*sample[2]/self.fx
ret[1] = (sample[1]-self.uy)*sample[2]/self.fy
ret[2] = sample[2]
return ret
########################################### MSRA ###################################
# Num_frame_per_each_subject = [ 8499, 8492, 8412, 8488, 8500, 8497, 8497, 8498, 8492]
class MSRAHandPoseDataset(HandPoseDataset):
def __init__(self, basepath="",train=True,cropSize=(128,128),doJitterRotation=False,doAddWhiteNoise=False,rotationAngleRange=[-45.0, 45.0],
comJitter=False,RandDotPercentage=0, indeces=None, sigmaNoise=1,cropSize3D=[250,250,250], do_norm_zero_one=False,random_seed=21,
drop_joint_num=0,center_refined=False,horizontal_flip=0, scale_aug=0, LeaveOut_subject=0 , use_default_cube=True):
self.fx, self.fy, self.ux , self.uy = (241.42, 241.42, 160., 120.)
self.num_joints = 21
self.default_cubes = {'P0': [200, 200, 200],
'P1': [200, 200, 200],
'P2': [200, 200, 200],
'P3': [180, 180, 180],
'P4': [180, 180, 180],
'P5': [180, 180, 180],
'P6': [170, 170, 170],
'P7': [160, 160, 160],
'P8': [150, 150, 150]}
self.address_gtUVD = os.path.join(basepath, "msra_test_groundtruth_label.txt")
self.address_files = os.path.join(basepath, "msra_test_list.txt")
self.use_default_cube = use_default_cube
data = {}
gt = open(self.address_gtUVD);gt.seek(0);
gt_lines = gt.readlines()
files = open(self.address_files);files.seek(0);
file_lines = files.readlines()
assert len(gt_lines) == len(file_lines)
for i in range(len(file_lines)):
part = gt_lines[i].split(' ')
gt_uvd_original = np.zeros((self.num_joints, 3), np.float32)
for joint in range(self.num_joints):
for xyz in range(0, 3):
gt_uvd_original[joint, xyz] = part[joint*3+xyz]
depth_address = os.path.join(basepath, file_lines[i][:-1])
subject = file_lines[i].split("/")[0]
data[subject] = data.get(subject,[]) + [ (gt_uvd_original, subject, depth_address) ]
if train:
self.seqName = "train"
# for key,value in data.items():
# if key != f"P{LeaveOut_subject}":
# print(key,len(data[key]))
# else:
# print(key,len(data[key])," **")
data = [data[key] for key in data.keys() if key != f"P{LeaveOut_subject}"]
self.data = [item for sublist in data for item in sublist]
else:
self.seqName = "test"
self.data=data[f"P{LeaveOut_subject}"]
self.numSamples = len(self.data)
super(MSRAHandPoseDataset, self).__init__(basepath=basepath,train=train,cropSize=cropSize,doJitterRotation=doJitterRotation,doAddWhiteNoise=doAddWhiteNoise,rotationAngleRange=rotationAngleRange,
comJitter=comJitter,RandDotPercentage=RandDotPercentage, indeces=indeces, sigmaNoise=sigmaNoise,cropSize3D=cropSize3D,
do_norm_zero_one=do_norm_zero_one,random_seed=random_seed,drop_joint_num=drop_joint_num,center_refined=False,horizontal_flip=horizontal_flip, scale_aug=scale_aug)
self.center_refined = center_refined
if center_refined:
print("Center refined being used")
self.center_refined_uvd = [i for i in range(self.numSamples)]
print("MSRA Dataset init done.")
def get_validIndex(self,ind):
return self.indeces[ind]
def LoadSample(self,ind,com=None):
idComGT = 9
# Load the dataset
gt_uvd_original, subject, depth_address = self.data[ind]
self.depth_address = depth_address
gt3Dorignal = self.joints3DToImg(gt_uvd_original)
gt2Dorignal = gt_uvd_original
numJoints = gt2Dorignal.shape[0]
dpt = self.loadDepthMap(depth_address).copy()
original_dpt = dpt.copy()
com = calculateCoM(dpt,minDepth=0,maxDepth=3000)
self.com=com
self.randomComJitter = np.clip(np.random.randn(3)*6,-self.comJitter,self.comJitter)# (1 if self.comJitter else 0)* np.clip(np.random.randn(3)*6,-6,6)
comGT = (copy.deepcopy(gt2Dorignal[idComGT]) if com is None else com) + self.randomComJitter #np.clip(np.concatenate((np.random.randn(2),np.array([0.])))*6,-6,6)
# Add noise?
if self.doAddWhiteNoise:
img_white_noise_scale = np.random.randn(dpt.shape[0], dpt.shape[1])
dpt = dpt + sigmaNoise * self.img_white_noise_scale
if self.doJitterRotation:
rotation_angle_scale = np.random.randn()
rot = rotation_angle_scale * (self.rotationAngleRange[1] - self.rotationAngleRange[0]) + self.rotationAngleRange[0]
self.randomAngle=rot
dpt, gt2Dorignal = rotateImageAndGt(dpt, comGT, rot, gt2Dorignal, bgValue=10000)
# Jitter scale (cube size)?
cubesize = (self.default_cubes[subject] if self.use_default_cube else self.cropSize3D)
dpt, M, com = self.cropArea3D(imgDepth=dpt,com=comGT,minRatioInside=0.6,size=cubesize, dsize=self.cropSize)
com3D = self.pointImgTo3D(com)
gt3Dcrop = gt3Dorignal - com3D # normalize to com
gt2Dcrop = np.zeros((gt2Dorignal.shape[0], 3), np.float32)
for joint in range(gt2Dorignal.shape[0]):
t=transformPoint2D(gt2Dorignal[joint], M)
gt2Dcrop[joint, 0] = t[0]
gt2Dcrop[joint, 1] = t[1]
gt2Dcrop[joint, 2] = gt2Dorignal[joint, 2]
D={};D["M"]=M;D["com3D"]=com3D;D["cubesize"]=cubesize
D["dpt"]=dpt.astype(np.float32);D["gt2Dorignal"]=gt2Dorignal;D["filename"]=depth_address
D["gt2Dcrop"]=gt2Dcrop;D["gt3Dorignal"]=gt3Dorignal;D["gt3Dcrop"]=gt3Dcrop;D["original_dpt"]=original_dpt;
return D
def convert_uvd_to_xyz_tensor( self, uvd ):
# uvd is a tensor of size(B,num_joints,3)
xyz = torch.zeros(uvd.shape);
xyz[:,:,2] = uvd[:,:,2];
xyz[:,:,0] = (uvd[:,:,0]-self.ux)*uvd[:,:,2]/self.fx
xyz[:,:,1] = (uvd[:,:,1]-self.uy)*uvd[:,:,2]/self.fy
return xyz
def convert_xyz_to_uvd_tensor(self, xyz):
# xyz is a tensor of size(B,num_joints,3)
uvd = torch.zeros(xyz.shape);
uvd[:,:,2] = xyz[:,:,2];
uvd[:,:,0] = xyz[:,:,0]/xyz[:,:,2]*self.fx+self.ux
uvd[:,:,1] = xyz[:,:,1]/xyz[:,:,2]*self.fy+self.uy
return uvd
def pointImgTo3D(self, sample):
"""
Normalize sample to metric 3D
:param sample: joints in (x,y,z) with x,y in image coordinates and z in mm
:return: normalized joints in mm
"""
ret = np.zeros((3,), np.float32)
# convert to metric using f
ret[0] = (sample[0]-self.ux)*sample[2]/self.fx
ret[1] = (sample[1]-self.uy)*sample[2]/self.fy
ret[2] = sample[2]
return ret
def joints3DToImg(self, sample):
"""
Denormalize sample from metric 3D to image coordinates
:param sample: joints in (x,y,z) with x,y and z in mm
:return: joints in (x,y,z) with x,y in image coordinates and z in mm
"""
ret = np.zeros((sample.shape[0], 3), np.float32)
for i in range(sample.shape[0]):
ret[i] = self.pointImgTo3D(sample[i])
return ret
def loadDepthMap(self, filename):
"""
Read a depth-map
:param filename: file name to load
:return: image data of depth image
"""
with open(filename, 'rb') as f:
# first 6 uint define the full image
width = struct.unpack('i', f.read(4))[0]
height = struct.unpack('i', f.read(4))[0]
left = struct.unpack('i', f.read(4))[0]
top = struct.unpack('i', f.read(4))[0]
right = struct.unpack('i', f.read(4))[0]
bottom = struct.unpack('i', f.read(4))[0]
patch = np.fromfile(f, dtype='float32', sep="")
imgdata = np.zeros((height, width), dtype='float32')
imgdata[top:bottom, left:right] = patch.reshape([bottom-top, right-left])
return imgdata
########################################### Functions ############################################################
def comToBounds(com, size, fx, fy):
"""
Calculate boundaries, project to 3D, then add offset and backproject to 2D (ux, uy are canceled)
:param com: center of mass, in image coordinates (x,y,z), z in mm
:param size: (x,y,z) extent of the source crop volume in mm
:return: xstart, xend, ystart, yend, zstart, zend
"""
zstart = com[2] - size[2] / 2.
zend = com[2] + size[2] / 2.
xstart = int(np.floor((com[0] * com[2] / fx - size[0] / 2.) / com[2]*fx+0.5))
xend = int(np.floor((com[0] * com[2] / fx + size[0] / 2.) / com[2]*fx+0.5))
ystart = int(np.floor((com[1] * com[2] / fy - size[1] / 2.) / com[2]*fy+0.5))
yend = int(np.floor((com[1] * com[2] / fy + size[1] / 2.) / com[2]*fy+0.5))
return xstart, xend, ystart, yend, zstart, zend
def rotateImageAndGt(imgDepth, center,angle, gtUvd,bgValue=10000):
"""
:param angle: rotation angle
:param center: a tuple (x,y), which is going to be the center of the rotation
from image coordinates to 3D coordinates
like transformations.pointsImgTo3D() (from the same file).
(To enable specific projections like for the NYU dataset)
"""
# Rotate image around given joint
center =(center[0], center[1])
rotationMat = cv2.getRotationMatrix2D(center, angle, 1.0)
sizeRotImg = (imgDepth.shape[1], imgDepth.shape[0])
imgRotated = cv2.warpAffine(src=imgDepth, M=rotationMat, dsize=sizeRotImg, flags=cv2.INTER_NEAREST, borderMode=cv2.BORDER_CONSTANT,
borderValue=bgValue)
# Rotate GT
gtUvd_ = gtUvd.copy()
gtUvdRotated = np.ones((gtUvd_.shape[0], 3), dtype=gtUvd.dtype)
gtUvdRotated[:,0:2] = gtUvd_[:,0:2]
gtUvRotated = np.dot(rotationMat, gtUvdRotated.T)
gtUvdRotated[:,0:2] = gtUvRotated.T
gtUvdRotated[:,2] = gtUvd_[:,2]
return imgRotated, gtUvdRotated
def get_visible(img , landmarks, cropSize=128, background_value=1,win_size=4):
mask = torch.zeros(landmarks.shape[0],1, dtype=torch.bool)
for j in range(landmarks.shape[0]):
x = np.int32(np.round(landmarks[j,0]))
y = np.int32(np.round(landmarks[j,1]))
if x>=0 and x<cropSize and y>=0 and y<cropSize:
left=max(0,x-win_size);right=min(x+win_size,cropSize)
bottom=max(0,y-win_size);top=min(y+win_size,cropSize)
window = img[0,bottom:top,left:right]
if torch.sum(window)/window.numel() < background_value-1e-6 :
mask[j,0]=True
return torch.from_numpy( np.float32(mask) )
def normalizeZeroOne(sample):
imgD = np.asarray(sample["dpt"].copy(), 'float32')
imgD[imgD == 0] = sample.com[2] + (sample['cubesize'][2] / 2.)
imgD -= (sample["com3D"][2] - (sample['cubesize'][2] / 2.))
imgD /= sample['cubesize'][2]
target = np.clip(np.asarray(sample["gt3Dcrop"], dtype='float32') / sample['cubesize'][2], -0.5, 0.5) + 0.5
return imgD, target
def normalizeMinusOneOne(sample):
imgD = np.asarray(sample["dpt"].copy(), 'float32')
imgD[imgD == 0] = sample["com3D"][2] + (sample['cubesize'][2] / 2.)
imgD -= sample["com3D"][2]
imgD /= (sample['cubesize'][2] / 2.)
target = np.clip(np.asarray(sample["gt3Dcrop"], dtype='float32')/ (sample['cubesize'][2] / 2.), -1, 1)
return imgD, target
def transformPoint2D(pt, M):
"""
Transform point in 2D coordinates
:param pt: point coordinates
:param M: transformation matrix
:return: transformed point
"""
pt2 = np.asmatrix(M.reshape((3, 3))) * np.matrix([pt[0], pt[1], 1]).T
return np.array([pt2[0] / pt2[2], pt2[1] / pt2[2]])
def loadDepthMap(filename):
"""
Read a depth-map
:param filename: file name to load
:return: image data of depth image
"""
with open(filename) as f:
img = Image.open(filename)
# top 8 bits of depth are packed into green channel and lower 8 bits into blue
assert len(img.getbands()) == 3
r, g, b = img.split()
r = np.asarray(r,np.int32)
g = np.asarray(g,np.int32)
b = np.asarray(b,np.int32)
dpt = np.bitwise_or(np.left_shift(g,8),b)
imgdata = np.asarray(dpt,np.float32)
return imgdata
def CropToOriginal(preds,matrices):
# preds is a tensor of shape (B,K,3)
# matrices is a tensor of shape (B,4,4), which is supposed to be the inverse matrix of each data sample
v=torch.cat([preds,torch.ones(preds.shape[0],preds.shape[1],1).to(preds.device)],dim=-1)
result=(v@matrices.transpose(-1,-2))
return result[:,:,:3]
def PixDropout(img,background_value,P,V=1):
# image is a tensor of size (1,H,W)
# background_value denotes the value using which the foreground pixels are computed
# P is the percentage of foreground pixels that are assigned the value V
# returns the same image with P percentage of its foregrounds pixels set to the value V
img=img.clone()
mask=abs(img[0]-background_value)>1e-6 # locate foreground pixels
y,x=torch.where(mask) # get their pixel coordinates
num_pix=np.int32(P*len(y))
indecies_toSet=np.random.choice(len(x),size=num_pix,replace=False)
img[0,y[indecies_toSet],x[indecies_toSet]]=V
return img
def Normalize_depth(preds,sizes,coms,add_com=False):
# preds is a tensor of shape (B,k,3)
# sizes is a tensor of shape (B,3)
#coms is a tensor of shape (B,3)
# this function denormalizes the depths of the prediction