forked from wbenbihi/hourglasstensorflow
-
Notifications
You must be signed in to change notification settings - Fork 0
/
predictClass.py
1291 lines (1220 loc) · 50 KB
/
predictClass.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
# -*- coding: utf-8 -*-
"""
Deep Human Pose Estimation
Project by Walid Benbihi
MSc Individual Project
Imperial College
Created on Mon Jul 17 15:50:43 2017
@author: Walid Benbihi
@mail : w.benbihi(at)gmail.com
@github : https://github.com/wbenbihi/hourglasstensorlfow/
Abstract:
This python code creates a Stacked Hourglass Model
(Credits : A.Newell et al.)
(Paper : https://arxiv.org/abs/1603.06937)
Code translated from 'anewell' github
Torch7(LUA) --> TensorFlow(PYTHON)
(Code : https://github.com/anewell/pose-hg-train)
Modification are made and explained in the report
Goal : Achieve Real Time detection (Webcam)
----- Modifications made to obtain faster results (trade off speed/accuracy)
This work is free of use, please cite the author if you use it!
"""
import sys
sys.path.append('./')
from hourglass_tiny import HourglassModel
from time import time, clock, sleep
import numpy as np
import tensorflow as tf
import scipy.io
from train_launcher import process_config
import cv2
#from yolo_tiny_net import YoloTinyNet
from yolo_net import YOLONet
from datagen import DataGenerator
import config as cfg
import threading
class PredictProcessor():
"""
PredictProcessor class: Give the tools to open and use a trained model for
prediction.
Dependency: OpenCV or PIL (OpenCV prefered)
Comments:
Every CAPITAL LETTER methods are to be modified with regard to your needs and dataset
"""
#-------------------------INITIALIZATION METHODS---------------------------
def __init__(self, config_dict):
""" Initializer
Args:
config_dict : config_dict
"""
self.params = config_dict
self.HG = HourglassModel(nFeat= self.params['nfeats'], nStack = self.params['nstacks'],
nModules = self.params['nmodules'], nLow = self.params['nlow'], outputDim = self.params['num_joints'],
batch_size = self.params['batch_size'], drop_rate = self.params['dropout_rate'], lear_rate = self.params['learning_rate'],
decay = self.params['learning_rate_decay'], decay_step = self.params['decay_step'], dataset = None, training = False,
w_summary = True, logdir_test = self.params['log_dir_test'],
logdir_train = self.params['log_dir_test'], tiny = self.params['tiny'],
modif = False, name = self.params['name'], attention = self.params['mcam'], w_loss=self.params['weighted_loss'] , joints= self.params['joint_list'])
self.graph = tf.Graph()
self.src = 0
self.cam_res = (480,640)
def color_palette(self):
""" Creates a color palette dictionnary
Drawing Purposes
You don't need to modify this function.
In case you need other colors, add BGR color code to the color list
and make sure to give it a name in the color_name list
/!\ Make sure those 2 lists have the same size
"""
#BGR COLOR CODE
self.color = [(241,242,224), (196,203,128), (136,150,0), (64,77,0),
(201,230,200), (132,199,129), (71,160,67), (32,94,27),
(130,224,255), (7,193,255), (0,160,255), (0,111,255),
(220,216,207), (174,164,144), (139,125,96), (100,90,69),
(252,229,179), (247,195,79), (229,155,3), (155,87,1),
(231,190,225), (200,104,186), (176,39,156), (162,31,123),
(210,205,255), (115,115,229), (80,83,239), (40,40,198)]
# Color Names
self.color_name = ['teal01', 'teal02', 'teal03', 'teal04',
'green01', 'green02', 'green03', 'green04',
'amber01', 'amber02', 'amber03', 'amber04',
'bluegrey01', 'bluegrey02', 'bluegrey03', 'bluegrey04',
'lightblue01', 'lightblue02', 'lightblue03', 'lightblue04',
'purple01', 'purple02', 'purple03', 'purple04',
'red01', 'red02', 'red03', 'red04']
self.classes_name = ["aeroplane", "bicycle", "bird",
"boat", "bottle", "bus", "car", "cat", "chair",
"cow", "diningtable", "dog", "horse", "motorbike",
"person", "pottedplant", "sheep",
"sofa", "train","tvmonitor"]
# Person ID = 14
self.color_class = [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19]
self.palette = {}
for i, name in enumerate(self.color_name):
self.palette[name] = self.color[i]
def LINKS_JOINTS(self):
""" Defines links to be joined for visualization
Drawing Purposes
You may need to modify this function
"""
self.links = {}
# Edit Links with your needed skeleton
LINKS = [(0,1),(1,2),(2,6),(6,3),(3,4),(4,5),(6,8),(8,13),(13,14),(14,15),(8,12),(12,11),(11,10)]
self.LINKS_ACP = [(0,1),(1,2),(3,4),(4,5),(7,8),(8,9),(10,11),(11,12)]
color_id = [1,2,3,3,2,1,5,27,26,25,27,26,25]
self.color_id_acp = [8,9,9,8,19,20,20,19]
# 13 joints version
#LINKS = [(0,1),(1,2),(2,3),(3,4),(4,5),(6,7),(6,11),(11,12),(12,13),(6,11),(10,9),(9,8)]
#color_id = [1,2,3,2,1,0,27,26,25,27,26,25]
# 10 lines version
# LINKS = [(0,1),(1,2),(3,4),(4,5),(6,8),(8,9),(13,14),(14,15),(12,11),(11,10)]
for i in range(len(LINKS)):
self.links[i] = {'link' : LINKS[i], 'color' : self.palette[self.color_name[color_id[i]]]}
# ----------------------------TOOLS----------------------------------------
def col2RGB(self, col):
"""
Args:
col : (int-tuple) Color code in BGR MODE
Returns
out : (int-tuple) Color code in RGB MODE
"""
return col[::-1]
def givePixel(self, link, joints):
""" Returns the pixel corresponding to a link
Args:
link : (int-tuple) Tuple of linked joints
joints : (array) Array of joints position shape = outDim x 2
Returns:
out : (tuple) Tuple of joints position
"""
return (joints[link[0]].astype(np.int), joints[link[1]].astype(np.int))
# ---------------------------MODEL METHODS---------------------------------
def model_init(self):
""" Initialize the Hourglass Model
"""
t = time()
with self.graph.as_default():
self.HG.generate_model()
print('Graph Generated in ', int(time() - t), ' sec.')
def load_model(self, load = None):
""" Load pretrained weights (See README)
Args:
load : File to load
"""
with self.graph.as_default():
self.HG.restore(load)
def _create_joint_tensor(self, tensor, name = 'joint_tensor',debug = False):
""" TensorFlow Computation of Joint Position
Args:
tensor : Prediction Tensor Shape [nbStack x 64 x 64 x outDim] or [64 x 64 x outDim]
name : name of the tensor
Returns:
out : Tensor of joints position
Comment:
Genuinely Agreeing this tensor is UGLY. If you don't trust me, look at
'prediction' node in TensorBoard.
In my defence, I implement it to compare computation times with numpy.
"""
with tf.name_scope(name):
shape = tensor.get_shape().as_list()
if debug:
print(shape)
if len(shape) == 3:
resh = tf.reshape(tensor[:,:,0], [-1])
elif len(shape) == 4:
resh = tf.reshape(tensor[-1,:,:,0], [-1])
if debug:
print(resh)
arg = tf.arg_max(resh,0)
if debug:
print(arg, arg.get_shape(), arg.get_shape().as_list())
joints = tf.expand_dims(tf.stack([arg // tf.to_int64(shape[1]), arg % tf.to_int64(shape[1])], axis = -1), axis = 0)
for i in range(1, shape[-1]):
if len(shape) == 3:
resh = tf.reshape(tensor[:,:,i], [-1])
elif len(shape) == 4:
resh = tf.reshape(tensor[-1,:,:,i], [-1])
arg = tf.arg_max(resh,0)
j = tf.expand_dims(tf.stack([arg // tf.to_int64(shape[1]), arg % tf.to_int64(shape[1])], axis = -1), axis = 0)
joints = tf.concat([joints, j], axis = 0)
return tf.identity(joints, name = 'joints')
def _create_prediction_tensor(self):
""" Create Tensor for prediction purposes
"""
with self.graph.as_default():
with tf.name_scope('prediction'):
self.HG.pred_sigmoid = tf.nn.sigmoid(self.HG.output[:,self.HG.nStack - 1], name= 'sigmoid_final_prediction')
self.HG.pred_final = self.HG.output[:,self.HG.nStack - 1]
self.HG.joint_tensor = self._create_joint_tensor(self.HG.output[0], name = 'joint_tensor')
self.HG.joint_tensor_final = self._create_joint_tensor(self.HG.output[0,-1] , name = 'joint_tensor_final')
print('Prediction Tensors Ready!')
#----------------------------PREDICTION METHODS----------------------------
def predict_coarse(self, img, debug = False, sess = None):
""" Given a 256 x 256 image, Returns prediction Tensor
This prediction method returns a non processed Output
Values not Bounded
Args:
img : Image -Shape (256 x256 x 3) -Type : float32
debug : (bool) True to output prediction time
Returns:
out : Array -Shape (nbStacks x 64 x 64 x outputDim) -Type : float32
"""
if debug:
t = time()
if img.shape == (256,256,3):
if sess is None:
out = self.HG.Session.run(self.HG.output, feed_dict={self.HG.img : np.expand_dims(img, axis = 0)})
else:
out = sess.run(self.HG.output, feed_dict={self.HG.img : np.expand_dims(img, axis = 0)})
else:
print('Image Size does not match placeholder shape')
raise Exception
if debug:
print('Pred: ', time() - t, ' sec.')
return out
def pred(self, img, debug = False, sess = None):
""" Given a 256 x 256 image, Returns prediction Tensor
This prediction method returns values in [0,1]
Use this method for inference
Args:
img : Image -Shape (256 x256 x 3) -Type : float32
debug : (bool) True to output prediction time
Returns:
out : Array -Shape (64 x 64 x outputDim) -Type : float32
"""
if debug:
t = time()
if img.shape == (256,256,3):
if sess is None:
out = self.HG.Session.run(self.HG.pred_sigmoid, feed_dict={self.HG.img : np.expand_dims(img, axis = 0)})
else:
out = sess.run(self.HG.pred_sigmoid, feed_dict={self.HG.img : np.expand_dims(img, axis = 0)})
else:
print('Image Size does not match placeholder shape')
raise Exception
if debug:
print('Pred: ', time() - t, ' sec.')
return out
def joints_pred(self, img, coord = 'hm', debug = False, sess = None):
""" Given an Image, Returns an array with joints position
Args:
img : Image -Shape (256 x 256 x 3) -Type : float32
coord : 'hm'/'img' Give pixel coordinates relative to heatMap('hm') or Image('img')
debug : (bool) True to output prediction time
Returns
out : Array -Shape(num_joints x 2) -Type : int
"""
if debug:
t = time()
if sess is None:
j1 = self.HG.Session.run(self.HG.joint_tensor, feed_dict = {self.HG.img: img})
else:
j1 = sess.run(self.HG.joint_tensor, feed_dict = {self.HG.img: img})
print('JT:', time() - t)
t = time()
if sess is None:
j2 = self.HG.Session.run(self.HG.joint_tensor_final, feed_dict = {self.HG.img: img})
else:
j2 = sess.run(self.HG.joint_tensor_final, feed_dict = {self.HG.img: img})
print('JTF:', time() - t)
if coord == 'hm':
return j1, j2
elif coord == 'img':
return j1 * self.params['img_size'] / self.params['hm_size'], j2 *self.params['img_size'] / self.params['hm_size']
else:
print("Error: 'coord' argument different of ['hm','img']")
else:
if sess is None:
j = self.HG.Session.run(self.HG.joint_tensor_final, feed_dict = {self.HG.img: img})
else:
j = sess.run(self.HG.joint_tensor_final, feed_dict = {self.HG.img: img})
if coord == 'hm':
return j
elif coord == 'img':
return j * self.params['img_size'] / self.params['hm_size']
else:
print("Error: 'coord' argument different of ['hm','img']")
def joints_pred_numpy(self, img, coord = 'hm', thresh = 0.2, sess = None):
""" Create Tensor for joint position prediction
NON TRAINABLE
TO CALL AFTER GENERATING GRAPH
Notes:
Not more efficient than Numpy, prefer Numpy for such operation!
"""
if sess is None:
hm = self.HG.Session.run(self.HG.pred_sigmoid , feed_dict = {self.HG.img: img})
else:
hm = sess.run(self.HG.pred_sigmoid , feed_dict = {self.HG.img: img})
joints = -1*np.ones(shape = (self.params['num_joints'], 2))
for i in range(self.params['num_joints']):
index = np.unravel_index(hm[0,:,:,i].argmax(), (self.params['hm_size'],self.params['hm_size']))
if hm[0,index[0], index[1],i] > thresh:
if coord == 'hm':
joints[i] = np.array(index)
elif coord == 'img':
joints[i] = np.array(index) * self.params['img_size'] / self.params['hm_size']
return joints
def batch_pred(self, batch, debug = False):
""" Given a 256 x 256 images, Returns prediction Tensor
This prediction method returns values in [0,1]
Use this method for inference
Args:
batch : Batch -Shape (batchSize x 256 x 256 x 3) -Type : float32
debug : (bool) True to output prediction time
Returns:
out : Array -Shape (batchSize x 64 x 64 x outputDim) -Type : float32
"""
if debug:
t = time()
if batch[0].shape == (256,256,3):
out = self.HG.Session.run(self.HG.pred_sigmoid, feed_dict={self.HG.img : batch})
else:
print('Image Size does not match placeholder shape')
raise Exception
if debug:
print('Pred: ', time() - t, ' sec.')
return out
#-------------------------------PLOT FUNCTION------------------------------
def plt_skeleton(self, img, tocopy = True, debug = False, sess = None):
""" Given an Image, returns Image with plotted limbs (TF VERSION)
Args:
img : Source Image shape = (256,256,3)
tocopy : (bool) False to write on source image / True to return a new array
debug : (bool) for testing puposes
sess : KEEP NONE
"""
joints = self.joints_pred(np.expand_dims(img, axis = 0), coord = 'img', debug = False, sess = sess)
if tocopy:
img = np.copy(img)
for i in range(len(self.links)):
position = self.givePixel(self.links[i]['link'],joints)
cv2.line(img, tuple(position[0])[::-1], tuple(position[1])[::-1], self.links[i]['color'][::-1], thickness = 2)
if tocopy:
return img
def plt_skeleton_numpy(self, img, tocopy = True, thresh = 0.2, sess = None, joint_plt = True):
""" Given an Image, returns Image with plotted limbs (NUMPY VERSION)
Args:
img : Source Image shape = (256,256,3)
tocopy : (bool) False to write on source image / True to return a new array
thresh : Joint Threshold
sess : KEEP NONE
joint_plt : (bool) True to plot joints (as circles)
"""
joints = self.joints_pred_numpy(np.expand_dims(img, axis = 0), coord = 'img', thresh = thresh, sess = sess)
if tocopy:
img = np.copy(img)*255
for i in range(len(self.links)):
l = self.links[i]['link']
good_link = True
for p in l:
if np.array_equal(joints[p], [-1,-1]):
good_link = False
if good_link:
position = self.givePixel(self.links[i]['link'],joints)
cv2.line(img, tuple(position[0])[::-1], tuple(position[1])[::-1], self.links[i]['color'][::-1], thickness = 2)
if joint_plt:
for p in range(len(joints)):
if not(np.array_equal(joints[p], [-1,-1])):
cv2.circle(img, (int(joints[p,1]), int(joints[p,0])), radius = 3, color = self.color[p][::-1], thickness = -1)
if tocopy:
return img
def pltSkeleton(self, img, thresh = 0.2, pltJ = True, pltL = True, tocopy = True, norm = True):
""" Plot skeleton on Image (Single Detection)
Args:
img : Input Image ( RGB IMAGE 256x256x3)
thresh : Joints Threshold
pltL : (bool) Plot Limbs
tocopy : (bool) Plot on imput image or return a copy
norm : (bool) Normalize input Image (DON'T MODIFY)
Returns:
img : Copy of input image if 'tocopy'
"""
if tocopy:
img = np.copy(img)
if norm:
img_hg = img / 255
hg = self.HG.Session.run(self.HG.pred_sigmoid, feed_dict = {self.HG.img: np.expand_dims(img_hg, axis = 0)})
j = np.ones(shape = (self.params['num_joints'],2)) * -1
for i in range(len(j)):
idx = np.unravel_index( hg[0,:,:,i].argmax(), (64,64))
if hg[0, idx[0], idx[1], i] > thresh:
j[i] = np.asarray(idx) * 256 / 64
if pltJ:
cv2.circle(img, center = tuple(j[i].astype(np.int))[::-1], radius= 5, color= self.color[i][::-1], thickness= -1)
if pltL:
for i in range(len(self.links)):
l = self.links[i]['link']
good_link = True
for p in l:
if np.array_equal(j[p], [-1,-1]):
good_link = False
if good_link:
pos = self.givePixel(l, j)
cv2.line(img, tuple(pos[0])[::-1], tuple(pos[1])[::-1], self.links[i]['color'][::-1], thickness = 5)
if tocopy:
return img
#------------------------Visualiazing Methods------------------------------
def jointsToMat(self, joints):
""" Given a 16 Joints Matrix, returns a 13 joints Matrix
MPII Formalism to PennAction DeepDragon Formalism
(Neck, Torso, Pelvis) erased
Args:
joints : Joints of shape (16,2) (HAS TO FOLLOW MPII FORMALISM)
Returns:
position : Joints of shape (13,2)
"""
position = np.zeros((13,2))
position[0:6,:] = joints[0:6,:]
position[6,:] = joints[9,:]
position[7:,:] = joints[10:,:]
position = np.reshape(position,(26,1),order = 'F')
return position
def computeErr(self, history, frame = 3):
""" Given frames, compute error vector to project into new basis
Args:
history : List of Joints detected in previous frames
frame : Number of previous frames to consider
Returns:
err : Error vector
"""
err = np.zeros((13*2*(frame-1),1))
rsho = 9
lhip = 3
eps = 0.00000001
for i in range(frame-1):
jf,js = history[-i-1], history[-i-2]
xf,yf = jf[:13], jf[13:]
xs,ys = js[:13], js[13:]
x = xf / abs(xf[rsho] - xf[lhip] + eps) - xs / abs(xs[rsho] - xs[lhip] + eps)
y = yf / abs(yf[rsho] - yf[lhip] + eps) - ys / abs(ys[rsho] - ys[lhip] + eps)
err[26*(i):26*(i)+26,:] = np.concatenate([x,y],axis = 0)
return err
def errToJoints(self, err, frame, hFrame):
""" Given an error vector and the frame considered, compute the joint's
location from the error
Args:
err : Error Vector
frame : Current Frame Detected Joints
hFrame : Previous Frames Detected Joints
Returns:
joints : Joints computed from error
"""
# Don't change Matrix computed with those joints to normalize
lhip = 3
rsho = 9
# Epsilon to avoir ZeroDivision Exception
eps = 0.00000001
# Compute X and Y coordinates
x = err[:,0:13]
y = err[:,13:26]
x = np.transpose(x)
y = np.transpose(y)
xh,yh = hFrame[0:13,:], hFrame[13:26,:]
xf,yf = frame[0:13,:], frame[13:26,:]
x = (x + (xh /np.abs(xh[rsho,:] - xh[lhip,:] + eps))) * np.abs(xf[rsho,:] - xf[lhip,:] + eps)
y = (y + (yh /np.abs(yh[rsho,:] - yh[lhip,:] + eps))) * np.abs(yf[rsho,:] - yf[lhip,:] + eps)
joints = np.concatenate([x,y], axis = 1).astype(np.int64)
return joints
def reconstructACPVideo(self, load = 'F:/Cours/DHPE/DHPE/hourglass_tiny/withyolo/p4frames.mat',n = 5):
""" Single Person Detection with Principle Componenent Analysis (PCA)
This method reconstructs joints given an error matrix (computed on MATLAB and available on GitHub)
and try to use temporal information from previous frames to improve detection and reduce False Positive
/!\Work In Progress on this model
/!\Our PCA considers the current frames and the last 3 frames
/!\WORKS ONLY ON 13 JOINTS FOR 16 JOINTS TRAINED MODELS
Args:
load : MATLAB file to load (the eigenvectors matrix should be called P for convenience)
n : Numbers of Dimensions to consider
"""
# OpenCv Video Capture : 0 is Webcam
cam = cv2.VideoCapture(self.src)
frame = 0
#/!\NOT USED YET
#rsho = 9
#lhip = 3
# Keep History of previous frames
jHistory = []
# Load Eigenvectors MATLAB matrix
P = scipy.io.loadmat(load)['P']
while True:
# Detection As Usual (See hpeWebcam)
t = time()
rec = np.zeros((500,500,3)).astype(np.uint8)
plt = np.zeros((500,500,3)).astype(np.uint8)
ret_val, img = cam.read()
img = cv2.flip(img,1)
img[:, self.cam_res[1]//2 - self.cam_res[0]//2:self.cam_res[1]//2 + self.cam_res[0]//2]
img_hg = cv2.resize(img, (256,256))
img_res = cv2.resize(img, (500,500))
img_hg = cv2.cvtColor(img_hg, cv2.COLOR_BGR2RGB)
j = self.HG.Session.run(self.HG.joint_tensor_final, feed_dict = {self.HG.img: np.expand_dims(img_hg/255, axis = 0)})
j = np.asarray(j * 500 / 64).astype(np.int)
joints = self.jointsToMat(j)
jHistory.append(joints)
# When enough frames are available
if frame > 4:
#Compute Error
err = np.transpose(self.computeErr(jHistory, frame = 4))
# Dismiss useless frame
jHistory.pop(0)
# Project Error into New Basis (PCA new basis)
projectedErr = np.dot(err, P)
nComp = n
# Reconstruct Error by dimensionality reduction
recErr = np.dot(projectedErr[:,:nComp], np.transpose(P[:,:nComp]))
# Compute joints position from reconstructed error
newJ = self.errToJoints(recErr, jHistory[-1], jHistory[-2])
for i in range(8):
#for i in [4,5,6,7]:
pos = self.givePixel(self.LINKS_ACP[i], newJ)
cv2.line(img_res, tuple(pos[0])[::-1], tuple(pos[1])[::-1], self.color[self.color_id_acp[i]][::-1], thickness = 8)
cv2.line(rec, tuple(pos[0])[::-1], tuple(pos[1])[::-1], self.color[self.color_id_acp[i]][::-1], thickness = 8)
for i in range(13):
#for i in [8,9,11,12]:
l = self.links[i]['link']
pos = self.givePixel(l, j)
cv2.line(img_res, tuple(pos[0])[::-1], tuple(pos[1])[::-1], self.links[i]['color'][::-1], thickness = 5)
cv2.line(plt, tuple(pos[0])[::-1], tuple(pos[1])[::-1], self.links[i]['color'][::-1], thickness = 5)
fps = 1/(time()-t)
cv2.putText(img_res, 'FPS: ' + str(fps)[:4], (60, 40), 2, 2, (0,0,0), thickness = 2)
toplot = np.concatenate([rec, img_res, plt], axis = 1)
cv2.imshow('stream', toplot)
frame += 1
if cv2.waitKey(1) == 27:
print('Stream Ended')
cv2.destroyAllWindows()
break
cv2.destroyAllWindows()
cam.release()
def _singleDetection(self, plt_j = True, plt_l = True):
""" /!\/!\DO NOT USE THIS FUNCTION/!\/!\
/!\/!\METHOD FOR TEST PURPOSES ONLY/!\/!\
PREFER HPE WEBCAM METHOD
"""
cam = cv2.VideoCapture(self.src)
frame = 0
position = np.zeros((32,1))
while True:
t = time()
ret_val, img = cam.read()
img = cv2.flip(img, 1)
img[:, self.cam_res[1]//2 - self.cam_res[0]//2:self.cam_res[1]//2 + self.cam_res[0]//2]
img_hg = cv2.resize(img, (256,256))
img_res = cv2.resize(img, (400,400))
cv2.imwrite('F:/Cours/DHPE/photos/acp/%04d.png' % (frame,), img_res)
img_hg = cv2.cvtColor(img_hg, cv2.COLOR_BGR2RGB)
j = self.HG.Session.run(self.HG.joint_tensor_final, feed_dict = {self.HG.img: np.expand_dims(img_hg/255, axis = 0)})
j = np.asarray(j * 400 / 64).astype(np.int)
joints = self.jointsToMat(j)
X = j.reshape((32,1),order = 'F')
position = np.hstack((position, X))
if plt_j:
for i in range(len(j)):
cv2.circle(img_res, center = tuple(j[i].astype(np.int))[::-1], radius= 5, color= self.color[i][::-1], thickness= -1)
if plt_l:
for i in range(len(self.links)):
l = self.links[i]['link']
pos = self.givePixel(l, j)
cv2.line(img_res, tuple(pos[0])[::-1], tuple(pos[1])[::-1], self.links[i]['color'][::-1], thickness = 5)
fps = 1/(time()-t)
cv2.putText(img_res, 'FPS: ' + str(fps)[:4], (60, 40), 2, 2, (0,0,0), thickness = 2)
cv2.imshow('stream', img_res)
frame += 1
if cv2.waitKey(1) == 27:
print('Stream Ended')
cv2.destroyAllWindows()
scipy.io.savemat('acpTest2.mat', dict(history = position))
break
cv2.destroyAllWindows()
cam.release()
def hpeWebcam(self, thresh = 0.6, plt_j = True, plt_l = True, plt_hm = False, debug = True):
""" Single Person Detector
Args:
thresh : Threshold for joint plotting
plt_j : (bool) To plot joints (as circles)
plt_l : (bool) To plot links/limbs (as lines)
plt_hm : (bool) To plot heat map
"""
cam = cv2.VideoCapture(self.src)
if debug:
framerate = []
while True:
t = time()
ret_val, img = cam.read()
img = cv2.flip(img, 1)
img[:, self.cam_res[1]//2 - self.cam_res[0]//2:self.cam_res[1]//2 + self.cam_res[0]//2]
img_hg = cv2.resize(img, (256,256))
img_res = cv2.resize(img, (800,800))
#img_copy = np.copy(img_res)
img_hg = cv2.cvtColor(img_hg, cv2.COLOR_BGR2RGB)
hg = self.HG.Session.run(self.HG.pred_sigmoid, feed_dict = {self.HG.img: np.expand_dims(img_hg/255, axis = 0)})
j = np.ones(shape = (self.params['num_joints'],2)) * -1
if plt_hm:
hm = np.sum(hg[0], axis = 2)
hm = np.repeat(np.expand_dims(hm, axis = 2), 3, axis = 2)
hm = cv2.resize(hm, (800,800))
img_res = img_res / 255 + hm
for i in range(len(j)):
idx = np.unravel_index( hg[0,:,:,i].argmax(), (64,64))
if hg[0, idx[0], idx[1], i] > thresh:
j[i] = np.asarray(idx) * 800 / 64
if plt_j:
cv2.circle(img_res, center = tuple(j[i].astype(np.int))[::-1], radius= 5, color= self.color[i][::-1], thickness= -1)
if plt_l:
for i in range(len(self.links)):
l = self.links[i]['link']
good_link = True
for p in l:
if np.array_equal(j[p], [-1,-1]):
good_link = False
if good_link:
pos = self.givePixel(l, j)
cv2.line(img_res, tuple(pos[0])[::-1], tuple(pos[1])[::-1], self.links[i]['color'][::-1], thickness = 5)
fps = 1/(time()-t)
if debug:
framerate.append(fps)
cv2.putText(img_res, 'FPS: ' + str(fps)[:4], (60, 40), 2, 2, (0,0,0), thickness = 2)
cv2.imshow('stream', img_res)
if cv2.waitKey(1) == 27:
print('Stream Ended')
cv2.destroyAllWindows()
break
cv2.destroyAllWindows()
cam.release()
if debug:
return framerate
def mpe(self, j_thresh = 0.5, nms_thresh = 0.5, plt_l = True, plt_j = True, plt_b = True, img_size = 800, skeleton = False):
""" Multiple Person Estimation (WebCam usage)
Args:
j_thresh : Joint Threshold
nms_thresh : Non Maxima Suppression Threshold
plt_l : (bool) Plot Limbs
plt_j : (bool) Plot Joints
plt_b : (bool) Plot Bounding Boxes
img_size : Resolution of Output Image
skeleton : (bool) Plot Skeleton alone next to image
"""
cam = cv2.VideoCapture(self.src)
res = img_size
fpsl = []
while True:
t = time()
if skeleton:
skel = np.zeros((img_size,img_size,3)).astype(np.uint8)
ret_val, img = cam.read()
img = cv2.flip(img,1)
img[:, self.cam_res[1]//2 - self.cam_res[0]//2:self.cam_res[1]//2 + self.cam_res[0]//2]
img_res = cv2.resize(img, (res,res))
img_yolo = np.copy(img_res)
img_yolo = cv2.cvtColor(img_yolo, cv2.COLOR_BGR2RGB)
results = self.detect(img_yolo)
results_person = []
for i in range(len(results)):
if results[i][0] == 'person':
results_person.append(results[i])
results_person = self.nms(results_person, nms_thresh)
for box in results_person:
class_name = box[0]
x = int(box[1])
y = int(box[2])
w = int(box[3] / 2)
h = int(box[4] / 2)
prob = box[5]
bbox = np.asarray((max(0,x-w), max(0, y-h), min(img_size-1, x+w), min(img_size-1, y+h)))
if plt_b:
cv2.rectangle(img_res, (bbox[0], bbox[1]), (bbox[2], bbox[3]), (0, 255, 0), 2)
cv2.rectangle(img_res, (bbox[0], bbox[1] - 20),(bbox[2], bbox[1]), (125, 125, 125), -1)
cv2.putText(img_res, class_name + ' : %.2f' % prob, (bbox[0] + 5, bbox[1] - 7), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0), 1)
maxl = np.max([w+0,h+0])
lenghtarray = np.array([maxl/h, maxl/w])
nbox = np.array([x-maxl, y-maxl, x+maxl, y+maxl])
padding = np.abs(nbox-bbox).astype(np.int)
img_person = np.copy(img_yolo[bbox[1]:bbox[3],bbox[0]:bbox[2],:])
padd = np.array([[padding[1],padding[3]],[padding[0],padding[2]],[0,0]])
img_person = np.pad(img_person, padd, mode = 'constant')
img_person = cv2.resize(img_person, (256,256))
hm = self.HG.Session.run(self.HG.pred_sigmoid, feed_dict={self.HG.img: np.expand_dims(img_person/255, axis= 0)})
j = -1*np.ones(shape = (self.params['num_joints'],2))
joint = -1*np.ones(shape = (self.params['num_joints'],2))
for i in range(self.params['num_joints']):
idx = np.unravel_index(hm[0,:,:,i].argmax(), (64,64))
if hm[0, idx[0], idx[1],i] > j_thresh:
j[i] = idx
joint[i] = np.asarray(np.array([y,x]) + ((j[i]-32)/32 * np.array([h,w])* lenghtarray )).astype(np.int)
if plt_j:
cv2.circle(img_res, tuple(joint[i].astype(np.int))[::-1], radius = 5, color = self.color[i][::-1], thickness = -1)
if plt_l:
for k in range(len(self.links)):
l = self.links[k]['link']
good_link = True
for p in l:
if np.array_equal(joint[p], [-1,-1]):
good_link = False
if good_link:
cv2.line(img_res, tuple(joint[l[0]][::-1].astype(np.int)), tuple(joint[l[1]][::-1].astype(np.int)), self.links[k]['color'][::-1], thickness = 3)
if skeleton:
cv2.line(skel, tuple(joint[l[0]][::-1].astype(np.int)), tuple(joint[l[1]][::-1].astype(np.int)), self.links[k]['color'][::-1], thickness = 3)
t_f = time()
cv2.putText(img_res, 'FPS: ' + str(1/(t_f-t))[:4], (60, 40), 2, 2, (0,0,0), thickness = 2)
fpsl.append(1/(t_f-t))
if skeleton:
img_res = np.concatenate([img_res,skel], axis = 1)
cv2.imshow('stream', img_res)
if cv2.waitKey(1) == 27:
print('Stream Ended')
cv2.destroyAllWindows()
return fpsl
break
cv2.destroyAllWindows()
cam.release()
# ---------------------------------MPE MULTITHREAD--------------------------
def threadProcessing(self, box, img_size, j_thresh, plt_l, plt_j, plt_b):
#if not coord.should_stop():
class_name = box[0]
x = int(box[1])
y = int(box[2])
w = int(box[3] / 2)
h = int(box[4] / 2)
prob = box[5]
bbox = np.asarray((max(0,x-w), max(0, y-h), min(img_size-1, x+w), min(img_size-1, y+h)))
if plt_b:
cv2.rectangle(self.img_res, (bbox[0], bbox[1]), (bbox[2], bbox[3]), (0, 255, 0), 2)
cv2.rectangle(self.img_res, (bbox[0], bbox[1] - 20),(bbox[2], bbox[1]), (125, 125, 125), -1)
cv2.putText(self.img_res, class_name + ' : %.2f' % prob, (bbox[0] + 5, bbox[1] - 7), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0), 1)
maxl = np.max([w+0,h+0])
lenghtarray = np.array([maxl/h, maxl/w])
nbox = np.array([x-maxl, y-maxl, x+maxl, y+maxl])
padding = np.abs(nbox-bbox).astype(np.int)
img_person = np.copy(self.img_yolo[bbox[1]:bbox[3],bbox[0]:bbox[2],:])
padd = np.array([[padding[1],padding[3]],[padding[0],padding[2]],[0,0]])
img_person = np.pad(img_person, padd, mode = 'constant')
img_person = cv2.resize(img_person, (256,256))
hm = self.HG.Session.run(self.HG.pred_sigmoid, feed_dict={self.HG.img: np.expand_dims(img_person/255, axis= 0)})
j = -1*np.ones(shape = (self.params['num_joints'],2))
joint = -1*np.ones(shape = (self.params['num_joints'],2))
for i in range(self.params['num_joints']):
idx = np.unravel_index(hm[0,:,:,i].argmax(), (64,64))
if hm[0, idx[0], idx[1],i] > j_thresh:
j[i] = idx
joint[i] = np.asarray(np.array([y,x]) + ((j[i]-32)/32 * np.array([h,w])* lenghtarray )).astype(np.int)
if plt_j:
cv2.circle(self.img_res, tuple(joint[i].astype(np.int))[::-1], radius = 5, color = self.color[i][::-1], thickness = -1)
if plt_l:
for k in range(len(self.links)):
l = self.links[k]['link']
good_link = True
for p in l:
if np.array_equal(joint[p], [-1,-1]):
good_link = False
if good_link:
cv2.line(self.img_res, tuple(joint[l[0]][::-1].astype(np.int)), tuple(joint[l[1]][::-1].astype(np.int)), self.links[k]['color'][::-1], thickness = 3)
# self.pprocessed += 1
# if self.pprocessed == self.pnum:
# coord.request_stop()
def imgPrepare(self, img, res):
img = cv2.flip(img,1)
img[:, self.cam_res[1]//2 - self.cam_res[0]//2:self.cam_res[1]//2 + self.cam_res[0]//2]
self.img_res = cv2.resize(img, (res,res))
self.img_yolo = np.copy(self.img_res)
self.img_yolo = cv2.cvtColor(self.img_yolo, cv2.COLOR_BGR2RGB)
def yoloPrepare(self, nms):
results = self.detect(self.img_yolo)
results_person = []
for i in range(len(results)):
if results[i][0] == 'person':
results_person.append(results[i])
results_person = self.nms(results_person, nms)
return results_person
def mpeThread(self, jThresh = 0.2, nms = 0.5, plt_l = True, plt_j = True, plt_b = True, img_size = 800, wait = 0.07):
cam = cv2.VideoCapture(self.src)
res = img_size
while True:
#coord = tf.train.Coordinator()
t = time()
ret_val, img = cam.read()
self.img_res, self.img_yolo = self.imgPrepare(img, res)
results_person = self.yoloPrepare(nms)
self.pnum = len(results_person)
self.pprocessed = 0
#workers = []
for box in results_person:
Thr = threading.Thread(target = self.threadProcessing, args = (box,img_size, jThresh, plt_l, plt_j, plt_b))
Thr.start()
#workers.append(Thr)
sleep(wait)
#coord.join(workers)
#for i in workers:
# i.join()
t_f = time()
cv2.putText(self.img_res, 'FPS: ' + str(1/(t_f-t))[:4], (60, 40), 2, 2, (0,0,0), thickness = 2)
cv2.imshow('stream', self.img_res)
if cv2.waitKey(1) == 27:
print('Stream Ended')
cv2.destroyAllWindows()
break
cv2.destroyAllWindows()
cam.release()
#---------------------------Conversion Method------------------------------
def videoDetection(self, src = None, outName = None, codec = 'DIVX', j_thresh = 0.5, nms_thresh = 0.5, show = True, plt_j = True, plt_l = True, plt_b = True):
""" Process Video with Pose Estimation
Args:
src : Source (video path or 0 for webcam)
outName : outName (set name of output file, set to None if you don't want to save)
codec : Codec to use for video compression (see OpenCV documentation)
j_thresh : Joint Threshold
nms_thresh : Non Maxima Suppression Threshold
show : (bool) True to show the video
plt_j : (bool) Plot Body Joints as circles
plt_l : (bool) Plot Limbs
plt_b : (bool) Plot Bounding Boxes
"""
cam = cv2.VideoCapture(src)
shape = np.asarray((cam.get(cv2.CAP_PROP_FRAME_HEIGHT),cam.get(cv2.CAP_PROP_FRAME_WIDTH))).astype(np.int)
frames = cam.get(cv2.CAP_PROP_FRAME_COUNT)
fps = cam.get(cv2.CAP_PROP_FPS)
if outName != None:
fourcc = cv2.VideoWriter_fourcc(*codec)
outVid = cv2.VideoWriter( outName, fourcc, fps, tuple(shape.astype(np.int))[::-1], 1)
cur_frame = 0
startT = time()
while (cur_frame < frames or frames == -1) and outVid.isOpened():
RECONSTRUCT_IMG = np.zeros((shape[0].astype(np.int),shape[1].astype(np.int),3))
ret_val, IMG_BASE = cam.read()
WIDTH = shape[1].astype(np.int)
HEIGHT = shape[0].astype(np.int)
XC = WIDTH // 2
YC = HEIGHT // 2
if WIDTH > HEIGHT:
top = np.copy(IMG_BASE[:,:XC - HEIGHT //2])
bottom = np.copy(IMG_BASE[:,XC + HEIGHT //2:])
img_square = np.copy(IMG_BASE[:,XC - HEIGHT //2:XC + HEIGHT //2])
elif HEIGHT > WIDTH:
top = np.copy(IMG_BASE[:YC - WIDTH //2])
bottom = np.copy(IMG_BASE[YC + WIDTH //2:])
img_square = np.copy(IMG_BASE[YC - WIDTH //2:YC + WIDTH //2])
else:
img_square = np.copy(IMG_BASE)
img_od = cv2.cvtColor(np.copy(img_square), cv2.COLOR_BGR2RGB)
shapeOd = img_od.shape
results = self.detect(img_od)
results_person = []
for i in range(len(results)):
if results[i][0] == 'person':
results_person.append(results[i])
results_person = self.nms(results_person, nms_thresh)
for box in results_person:
class_name = box[0]
x = int(box[1])
y = int(box[2])
w = int(box[3] / 2)
h = int(box[4] / 2)
prob = box[5]
bbox = np.asarray((max(0,x-w), max(0, y-h), min(shapeOd[1]-1, x+w), min(shapeOd[0]-1, y+h)))
if plt_b:
cv2.rectangle(img_square, (bbox[0], bbox[1]), (bbox[2], bbox[3]), (0, 255, 0), 2)
cv2.rectangle(img_square, (bbox[0], bbox[1] - 20),(bbox[2], bbox[1]), (125, 125, 125), -1)
cv2.putText(img_square, class_name + ' : %.2f' % prob, (bbox[0] + 5, bbox[1] - 7), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0), 1)
maxl = np.max([w+0,h+0])
lenghtarray = np.array([maxl/h, maxl/w])
nbox = np.array([x-maxl, y-maxl, x+maxl, y+maxl])
padding = np.abs(nbox-bbox).astype(np.int)
img_person = np.copy(img_od[bbox[1]:bbox[3],bbox[0]:bbox[2],:])
padd = np.array([[padding[1],padding[3]],[padding[0],padding[2]],[0,0]])
img_person = np.pad(img_person, padd, mode = 'constant')
img_person = cv2.resize(img_person, (256,256))
hm = self.HG.Session.run(self.HG.pred_sigmoid, feed_dict={self.HG.img: np.expand_dims(img_person/255, axis= 0)})
j = -1*np.ones(shape = (self.params['num_joints'],2))
joint = -1*np.ones(shape = (self.params['num_joints'],2))
for i in range(self.params['num_joints']):
idx = np.unravel_index(hm[0,:,:,i].argmax(), (64,64))
if hm[0, idx[0], idx[1],i] > j_thresh:
j[i] = idx
joint[i] = np.asarray(np.array([y,x]) + ((j[i]-32)/32 * np.array([h,w])* lenghtarray )).astype(np.int)
if plt_j:
cv2.circle(img_square, tuple(joint[i].astype(np.int))[::-1], radius = 5, color = self.color[i][::-1], thickness = -1)
if plt_l:
for k in range(len(self.links)):
l = self.links[k]['link']
good_link = True
for p in l:
if np.array_equal(joint[p], [-1,-1]):
good_link = False
if good_link:
cv2.line(img_square, tuple(joint[l[0]][::-1].astype(np.int)), tuple(joint[l[1]][::-1].astype(np.int)), self.links[k]['color'][::-1], thickness = 3)
if WIDTH > HEIGHT:
RECONSTRUCT_IMG[:,:XC - HEIGHT //2] = top
RECONSTRUCT_IMG[:,XC + HEIGHT //2:] = bottom
RECONSTRUCT_IMG[:,XC - HEIGHT //2: XC + HEIGHT //2] = img_square
elif HEIGHT < WIDTH:
RECONSTRUCT_IMG[:YC - WIDTH //2] = top
RECONSTRUCT_IMG[YC + WIDTH //2:] = bottom
RECONSTRUCT_IMG[YC - WIDTH //2:YC + WIDTH //2] = img_square
else:
RECONSTRUCT_IMG = img_square.astype(np.uint8)
RECONSTRUCT_IMG = RECONSTRUCT_IMG.astype(np.uint8)
outVid.write(np.uint8(RECONSTRUCT_IMG))
cur_frame = cur_frame + 1
if frames != -1:
percent = ((cur_frame+1)/frames) * 100
num = np.int(20*percent/100)
tToEpoch = int((time() - startT) * (100 - percent)/(percent))
sys.stdout.write('\r Processing: {0}>'.format("="*num) + "{0}>".format(" "*(20-num)) + '||' + str(percent)[:4] + '%' + ' -timeToEnd: ' + str(tToEpoch) + ' sec.')
sys.stdout.flush()
if show:
cv2.imshow('stream', RECONSTRUCT_IMG)
if cv2.waitKey(1) == 27:
print('Stream Ended')
cv2.destroyAllWindows()
cam.release()
print(time() - startT)
#if outName != None:
#outVid.release()
cv2.destroyAllWindows()
cam.release()
if outName != None:
print(outVid.isOpened())
outVid.release()
print(outVid.isOpened())
print(time() - startT)
#-------------------------Benchmark Methods (PCK)-------------------------
def pcki(self, joint_id, gtJ, prJ, idlh = 3, idrs = 12):
""" Compute PCK accuracy on a given joint
Args:
joint_id : Index of the joint considered
gtJ : Ground Truth Joint
prJ : Predicted Joint
idlh : Index of Normalizer (Left Hip on PCK, neck on PCKh)
idrs : Index of Normalizer (Right Shoulder on PCK, top head on PCKh)
Returns:
(float) NORMALIZED L2 ERROR
"""
return np.linalg.norm(gtJ[joint_id]-prJ[joint_id][::-1]) / np.linalg.norm(gtJ[idlh]-gtJ[idrs])
def pck(self, weight, gtJ, prJ, gtJFull, boxL, idlh = 3, idrs = 12):
""" Compute PCK accuracy for a sample
Args:
weight : Index of the joint considered
gtJFull : Ground Truth (sampled on whole image)
gtJ : Ground Truth (sampled on reduced image)
prJ : Prediction
boxL : Box Lenght
idlh : Index of Normalizer (Left Hip on PCK, neck on PCKh)
idrs : Index of Normalizer (Right Shoulder on PCK, top head on PCKh)