-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy pathindex.html
More file actions
1044 lines (824 loc) · 43.6 KB
/
index.html
File metadata and controls
1044 lines (824 loc) · 43.6 KB
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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta http-equiv="X-UA-Compatible" content="IE=edge">
<meta name="viewport" content="width=device-width, initial-scale=1">
<!-- The above 3 meta tags *must* come first in the head; any other head content must come *after* these tags -->
<link rel="shortcut icon" type="image/x-icon" href="triangle.ico">
<title>X-TC | Cross-Task Consistency</title>
<meta name="keywords" content="consistency, robust-learning, multitask, computer vision, vision, stanford, tasks, berkeley, EPFL, Swiss Federal Institute of Technology, Lausanne">
<meta name="description" content="Robust Learning Through Cross-Task Consistency EPFL, Stanford, UC Berkeley. We propose a flexible and fully computational framework for learning, while respecting consistency across an arbitrary dictionary of tasks.">
<!--meta name="google-site-verification" content="tBb4aimY-zxYS_s_z86g6faxMX2G_I_Wp_kzbF3U_RM" />--
<!-- Google font -->
<link href="https://fonts.googleapis.com/css?family=Montserrat:400,700%7CVarela+Round" rel="stylesheet">
<!-- Bootstrap -->
<link type="text/css" rel="stylesheet" href="css/bootstrap.min.css" />
<!-- Owl Carousel -->
<link type="text/css" rel="stylesheet" href="css/owl.carousel.css" />
<link type="text/css" rel="stylesheet" href="css/owl.theme.default.css" />
<!-- Magnific Popup -->
<link type="text/css" rel="stylesheet" href="css/magnific-popup.css" />
<!-- Icons -->
<link rel="stylesheet" href="css/font-awesome.min.css">
<link rel="stylesheet" href="https://cdn.rawgit.com/jpswalsh/academicons/master/css/academicons.min.css">
<!-- Custom stlylesheet -->
<link type="text/css" rel="stylesheet" href="css/style.css" />
<link type="text/css" rel="stylesheet" href="css/index.css" />
<link type="text/css" rel="stylesheet" href="css/cvpr-video.css"/>
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.2.1/jquery.min.js"></script>
<script src="//code.jquery.com/jquery-3.2.1.min.js"></script>
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/fancybox/3.2.5/jquery.fancybox.min.css" />
<script src="https://cdnjs.cloudflare.com/ajax/libs/fancybox/3.2.5/jquery.fancybox.min.js"></script>
<link rel="shortcut icon" type="image/x-icon" href="triangle.ico">
<!-- HTML5 shim and Respond.js for IE8 support of HTML5 elements and media queries -->
<!-- WARNING: Respond.js doesn't work if you view the page via file:// -->
<!--[if lt IE 9]>
<script src="https://oss.maxcdn.com/html5shiv/3.7.3/html5shiv.min.js"></script>
<script src="https://oss.maxcdn.com/respond/1.4.2/respond.min.js"></script>
<![endif]-->
<!-- Global site tag (gtag.js) - Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=UA-89674082-5"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'UA-89674082-5');
</script>
</head>
<body>
<!-- Header -->
<header id="home">
<!-- Background Image -->
<div style="" id='particles-js'>
</div>
<!-- /Background Image -->
<!-- Nav -->
<nav id="nav" class="navbar nav-transparent">
<div class="container">
<div class="navbar-header">
<!-- Logo -->
<div class="navbar-brand">
<a class="logo-text" href="//consistency.epfl.ch">X-TC</a>
<a href="http://vilab.epfl.ch/">
<img class="logo" src="" alt="logo">
<img class="logo-alt" src="" alt="logo">
</a>
<!--a href="http://bair.berkeley.edu/">
<img class="logo" src="./img/berkeley-logo.png" alt="logo">
<img class="logo-alt" src="./img/berkeley-logo-white.png" alt="logo">
<!-- <img class="logo-alt" src="./img/logo-alt.png" alt="logo"> -->
<!--/a-->
</div>
<!-- /Logo -->
<!-- Collapse nav button -->
<div class="nav-collapse">
<span></span>
</div>
<!-- /Collapse nav button -->
</div>
<!-- Main navigation -->
<ul class="main-nav nav navbar-nav navbar-right">
<li><a href="#home">Home</a></li>
<li><a href="./demo">Live Demo</a></li>
<li><a href="./visuals">Visualization</a></li>
<li><a href="#models">Pretrained Models</a></li>
<li><a href="#paper">Paper</a></li>
<li><a href="https://github.com/EPFL-VILAB/XTConsistency">Code</a></li>
<li><a href="#team">Team</a></li>
<!-- <li class="has-dropdown"><a href="#blog">Blog</a> -->
<!-- <ul class="dropdown"> -->
<!-- <li><a href="blog-single.html">blog post</a></li> -->
<!-- </ul> -->
<!-- </li> -->
<!-- <li><a href="#contact">Contact</a></li> -->
</ul>
<!-- /Main navigation -->
</div>
</nav>
<!-- /Nav -->
<!-- home wrapper -->
<div class="home-wrapper">
<div class="container">
<div class="row">
<!-- home content -->
<div class="col-md-10 col-md-offset-1">
<div class="home-content">
<h2 class="white-text">Robust Learning Through Cross-Task Consistency</h2><br>
<h4 class="white-text">CVPR 2020 [Best Paper Award Nominee]</h4>
<!--h2 class="white-text">(<u>X-TAC</u>)</h2-->
<!--<h3 class="white-text">(Under Development. Please check after March 12 for live demo & final content)</h3> -->
<a class="white-text" href="http://www.cs.stanford.edu/~amirz/">Zamir*</a><span style="opacity:0.7">,</span>
<a class="white-text" href="https://alexsax.github.io/">Sax*</a><span style="opacity:0.7">,</span>
<a class="white-text" href="https://aserety.github.io/">Yeo</a><span style="opacity:0.7">,</span>
<a class="white-text" href="https://ofkar.github.io/">Kar</a><span style="opacity:0.7">,</span>
<a class="white-text" href="#team">Cheerla</a><span style="opacity:0.7">,</span>
<br>
<a class="white-text" href="#team">Suri</a><span style="opacity:0.7">,</span>
<a class="white-text" href="#team">Cao</a><span style="opacity:0.7">,</span>
<a class="white-text" href="https://people.eecs.berkeley.edu/~malik/">Malik</a><span style="opacity:0.7">,</span>
<a class="white-text" href="https://geometry.stanford.edu/member/guibas/">Guibas</a><span style="opacity:0.7"></span>
<br>
<br>
<br>
<button id="learnmore" class="outline-btn roundedbutton">Learn more</button>
</div>
</div>
<!-- /home content -->
</div>
</div>
</div>
<!-- /home wrapper -->
</header>
<!-- /Header -->
<!-- Teaser -->
<div id="teaser" class="section bg-grey">
<!-- Container -->
<div class="container">
<div class="row" style="margin: 0px">
<!-- service -->
<div class="col-md-12 col-sm-6">
<div class="service">
<div class="service-content service-row-2">
<h3>Live Demo</h3>
<p>
Upload an image to see results of learning with Cross-Task Consistency versus various baselines for different vision tasks.</p>
<br>
<a href="./demo"> <img width="6%" src="./img/icon-demo1.png" > </a>
</div>
</div>
</div>
<!-- /service -->
</div>
<!-- Row -->
<div class="row" style="margin: 0px">
<!-- service -->
<div class="col-md-3 col-sm-6">
<div class="service smoothscroll" href="#paper">
<div class="service-content service-row-1">
<i class="fa fa-file-text-o"></i>
<h3>Paper</h3>
<p>The paper and supplementary material.</p>
<a href="#paper">PDF</a>
</div>
</div>
</div>
<!-- /service -->
<!-- service -->
<div class="col-md-3 col-sm-6">
<div class="service smoothscroll" href="#transfers" onclick="window.location.href='./visuals'">
<div class="service-content service-row-1">
<i class="fa fa-video-camera"></i>
<h3>Visualization</h3>
<p>Examine the visualization of consistencies across prediction tasks.</p>
<a href="./visuals">Visualization Page</a>
</div>
</div>
</div>
<!-- /service -->
<!-- service -->
<div class="col-md-3 col-sm-6">
<div class="service smoothscroll" href="#models">
<div class="service-content service-row-1">
<i class="fa fa-cogs"></i>
<h3>Models</h3>
<p>Download the trained consistency models and baselines.</p>
<a href="#models">Pretrained Models</a>
</div>
</div>
</div>
<!-- /service -->
<!-- service -->
<div class="col-md-3 col-sm-6">
<div class="service smoothscroll" href="https://github.com/EPFL-VILAB/XTConsistency">
<div class="service-content service-row-1">
<i class="fa fa-github"></i>
<h3>Code</h3>
<p>The code for training and testing consistency models and baselines.</p>
<a href="https://github.com/EPFL-VILAB/XTConsistency">Get started</a>
</div>
</div>
</div>
<!-- /service -->
</div>
<!-- /Row -->
<!-- /Row -->
</div>
<!-- /Container -->
</div>
<!-- /Teaser -->
<!-- Abstract -->
<div id="abstract" class="section sm-padding">
<!-- Container -->
<div class="container">
<!-- Row -->
<div class="row">
<!-- Section header -->
<div class="section-header text-center">
<h2 class="title">Quick Summary</h2>
<u><strong>What</strong></u> is consistency, <u><strong>why</strong></u> is it important, and <u><strong>how</strong></u> do we use it?
</div>
<!-- /Section header -->
<div class="abstract-body col-md-12">
<p>
<div class="figure row">
<img width="80%" src="./img_consistency/pull_april22_3.jpg" alt="" class="center">
<p class='col-sm-10 col-sm-offset-1' style='text-align: justify;'>
<strong>Cross-Task Consistent Learning.</strong>
The lower and upper rows show the results of the baseline (independent learning) and learning with cross-task consistency. The latter yields higher quality (especially at hard-to-predict fine-grained details) and more consistent predictions. Red boxes provide magnifications.
</p>
</div>
<br>
<strong>1) <u><b>What</b> is consistency?</u> </strong>
<br>
Suppose an object detector detects a ball in a particular region of an image, while a depth estimator returns a flat surface for the same region. This presents an issue -- at least one of them has to be wrong, because they are <strong>inconsistent</strong>. More concretely, the first prediction domain (objects) and the second prediction domain (depth) are not independent and consequently enforce some constraints on each other, often referred to as <strong>consistency constraints</strong>.
<br>
<br>
<strong>2) <u> <b>Why</b> is it important to consider consistency in learning? </u> </strong>
<br>
First, desired learning tasks are usually predictions of different aspects of one underlying reality (the scene that underlies an image). Hence inconsistency among predictions implies contradiction and is inherently undesirable. Second, consistency constraints are informative and can be used to better fit the data or lower the sample complexity. Also, they may reduce the tendency of neural networks to learn "surface statistics'' (superficial cues), by enforcing constraints rooted in different physical or geometric rules. This is empirically supported by the improved generalization of models when trained with consistency constraints.
<br>
<br>
The video below demonstrates the impact of disregarding consistency in learning as well as the effectiveness of augmenting learning with cross-task consistency constraints. Each window shows surface normals predicted out of a domain itself predicted out of the image (i.e. image→{prediction domain X}→surface normals). The normals in the upper row (learning without consistency) are poor and different/inconsistent with each other for the same underlying image. The lower row shows the same except when learning image→{prediction domain X} was augmented with cross-task consistency constraints with normals. Inferred surface normals look better and more similar to each other regardless of the middle prediction domain, which demonstrates all of the middle domains were successfully made cross-task consistent w.r.t to normals. In the paper, we further extend this concept to many arbitrary domains with arbitrary inference path lengths, using a general and fully computational learning framework.
</p>
<div class="figure" style="text-align:center;">
<h4 class="">(Note: Videos on page best seen in HD)</h4>
<iframe width="1022" height="359" src="https://www.youtube.com/embed/W4F1Q_MM2Rw?rel=0&controls=0&showinfo=0&loop=1&playlist=W4F1Q_MM2Rw&modestbranding=1&autoplay=1&mute=1&start=0&vq=hd1080" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen=""></iframe>
</div>
<p>
<br>
<br>
<strong>3) <u> <b>How</b> can we design a learning system that makes consistent predictions? </u> </strong>
<br>
This paper proposes a method which, given an arbitrary dictionary of tasks, augments the learning objective with explicit constraints for cross-task consistency. The constraints are learned from data rather than apriori given relationships. For instance, it is not necessary to encode that surface normals are the 3D derivative of depth or occlusion edges are discontinuities in depth. This makes the method applicable to any pairs of tasks as long as they are not statistically independent; even if their analytical relationship is unknown, hard to program, or non-differentiable.
<br>
<br>
The primary concept behind the method is <emph>inference-path invariance</emph>. That is, the result of inferring an output domain from an input domain should be the same, regardless of the intermediate domains mediating the inference. When inference paths with the same endpoints, but different intermediate domains, yield similar results, this implies the intermediate domain predictions did not conflict as far as the output was concerned. We apply this concept over paths in a graph of tasks, where the nodes and edges are prediction domains and neural network mappings between them, respectively. Satisfying this invariance constraint over <emph>all</emph> paths in the graph ensures the predictions for all domains are in global cross-task agreement.
<br>
<br>
</p>
<div class="figure row">
<img width="80%" src="./img_consistency/graph_overview_april15_1.jpg" alt="" class="center">
<p class='col-sm-10 col-sm-offset-1' style='text-align: justify;'><strong>Enforcing cross-task consistency.</strong> <strong>(a)</strong> shows the typical multitask setup where predictors x→y1 and x→y2 are trained without a notation of consistency (either completely independently or with a shared encoder and dedicated decoders). <strong>(b)</strong> depicts the elementary <strong>triangle consistency constraint</strong> where the prediction x→y1 is enforced to be consistent with x→y2 using a function that relates y1 to y2 (i.e. y1→y2). <strong>(c)</strong> shows how the triangle unit from (b) can be an element of a larger system of domains. Finally, <strong>(d)</strong> illustrates the generalized case where in the larger system of domains, consistency can be enforced using invariance along arbitrary paths, as long as their endpoints are the same (here the blue and green paths). This is the general concept behind "inference-path invariance". The triangle in (b) is the smallest unit of such paths.
</div>
The figure below shows the effect of training with and without cross-task consistency, for networks trained to do surface normal predictions:
<br>
<br>
<div class="figure row">
<img width="80%" src="./img_consistency/fig_method_results_april17_1.jpg" alt="" class="center">
<p class='col-sm-10 col-sm-offset-1' style='text-align: justify;'><strong>Learning with and without cross-task consistency shown for a sample query.</strong> Using the notation x→y1→Y, here x is RGB image, y1 is surface normals, and five domains in Y are reshading, 3D curvature, texture edges (Sobel filter), depth, and occlusion edges.
<strong>Top row</strong> shows the results of standard training of x→y1. After convergence of training, the predicted normals (y1) are projected onto other domains (Y) which reveal various inaccuracies. This demonstrates such cross-task projections y1→Y can provide additional cues to training x→y1.
<strong>Middle row</strong> shows the results of consistent training of x→y1 by leveraging y1→Y in the loss. The predicted normals are notably improved, especially in hard to predict <strong>fine-grained details</strong> (zoom into the yellow markers).
<strong>Bottom row</strong> provides the ground truth.
</p>
</div>
<div class="figure" style="text-align:center">
<h4 class="">(Note: Video best seen in HD)</h4>
<iframe width="1022" height="311" src="https://www.youtube.com/embed/SywH7IbKn9E?rel=0&controls=0&showinfo=0&loop=1&playlist=SywH7IbKn9E&modestbranding=1&autoplay=1&mute=1&start=0&vq=hd1080" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen=""></iframe>
<p> This video demonstrates the same results as the above figure, but over a video and for depth prediction. The video is processed <u>frame-by-frame</u>.</p>
</div>
</div>
</div>
<!-- /Row -->
</div>
<!-- /Container -->
</div>
<!-- /Summary -->
<!-- Tryit -->
<div id="try_it" class="section md-padding bg-grey">
<!-- Container -->
<div class="container">
<!-- Row -->
<div class="row">
<!-- Section header -->
<div class="section-header text-center">
<h2 class="title">Who can use cross-task consistency?</h2>
<u><strong>Cross-Task constraints are useful for common single-task learning too, not just multi-task.</strong></u> Here's how.
</div>
<div class='row'>
<div class="col-sm-6">
<p>
Let's say that you're interested in predicting some y1 given x (For example surface normals given an RGB image).
Instead of simply training a network to do x→y1 (i.e. RGB→Surface Normals), train it to do x→y1→y2 (i.e. RGB→Surface Normals→y2). It will fit the data better with improved RGB→Surface Normal predictions.
</p>
</div>
<div class="col-sm-6 figure pull-right" style="text-align:center; margin-top:0px;">
<img width="100%" src="./assets/main_page/xtc_for_single_tasks.png" alt="How to use cross-task consistency for single-task learning." class="center">
<!-- <p style='text-align: justify;'><strong>Cross-task consistency can augment single-task learning.</strong> Instead of using independent learning, add consistency constraints from additional tasks or pretrained networks.</p> -->
</div>
</div>
<br>
<p>
Many options for y2 work; even if you have only (x,y1) training data. All choices gave some benefit, though some more than others. Some y2 options:
<ul style="list-style-type: circle;">
<li>
<strong>Use tasks that need no supervision as y2.</strong> For instance, 2D texture edges (sobel) or 2D keypoints (SURF). The figure below shows that adding such consistency constraints improved the RGB→Surface Normals results. See section "Using Consistency with Unsupervised Tasks" in the paper.
</li>
<li>
<strong>Use our released pre-trained cross-task functions</strong> as y1→y2 in your training job. Those models can be downloaded from our <a href='https://github.com/EPFL-VILAB/XTConsistency#pretrained-models'>github repository</a>.
</li>
</ul>
</p>
<div class="figure row" style="text-align:center">
<div class='col-sm-10 col-sm-offset-1'>
<img width="60%" src="./assets/main_page/unsupervised_results.jpg" alt="Results from using cross-task consistency with unsupervised tasks for y2." class="center">
<p style='text-align: justify;'><strong>Results from using unsupervised consistency constraints.</strong> The network trained with cross-task consistency used only 2D edges and 2D keypoints as 'y2' tasks. Since these are unsupervised tasks, they required no additional supervision. Still, the consistency-trained network better fits the fine-grained details in images, as shown in the area highlighted with yellow circles. This echoes cross-task constraints are useful for common single-task learning too, even when no additional tasks/data/labels are available.</p>
</div>
</div>
To summarize: cross-task consistency can be used to improve most every-day neural network training jobs, not just multi-task learning.
</div>
<!-- /Row -->
</div>
<!-- /Tryit -->
</div>
<!-- /Summary -->
<!-- Sample Results -->
<div id="sample" class="section sm-padding">
<!-- Container -->
<div class="container">
<!-- Row -->
<div class="row">
<!-- Section header -->
<div class="section-header text-center">
<h2 class="title">Sample Results</h2>
</div>
<!-- /Section header -->
<div class="abstract-body col-md-12">
<p>
Below you can see sample results of learning with cross-task consistency for three sample prediction domains (surface normals, depth, (re)shading) over video and image queries. The video results are on external YouTube videos and the predictions are made <u>frame-by-frame</u> with no temporal smoothing. <u>Zoom in to see the fine-grained details</u>.
</p>
<div class="figure row">
<img width="80%" src="./img_consistency/domain_comparison_april21_1.jpg" alt="" class="center">
<p class='col-sm-10 col-sm-offset-1' style='text-align: justify;'>
<strong>Cross-Task Consistent learning results on external domain data.</strong> Queries: Bedroom in Arles, Van Gogh (1888); Cotton Mill Girl, Lewis Hine (1908); Chernobyl Pripyat Abandoned School (c. 2009). Try the models on your own images at the <a href="./demo">live demo page</a>.
</p>
</div>
<div class="figure" style="text-align:center">
<h4 class="">(Note: Video best seen in HD)</h4>
<iframe width="948" height="721" src="https://www.youtube.com/embed/6CSmmrBNX9M?rel=0&controls=0&showinfo=0&loop=1&playlist=6CSmmrBNX9M&modestbranding=1&autoplay=1&mute=1&start=1&vq=hd1080" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen=""></iframe>
</div>
<br>
<br>
<br>
<p> <strong>Comparison with baselines:</strong> Below are some representative results from using cross-task consistent learning and various baselines. You can run comparisons yourself by uploading images on our <a href="./demo">live demo</i></a>
</p>
<div class="figure row">
<img width="80%" src="./img_consistency/method_comparison_april20_2.jpg" alt="" class="center">
<p class='col-sm-10 col-sm-offset-1' style='text-align: justify;'>
<strong>Networks trained with cross-task consistency, or with alternative baselines.</strong> The queries are from Taskonomy dataset <strong>(top)</strong> or external data <strong>(Cotton Mill Girl, Lewis Hine 1908; bottom)</strong>.
</p>
</div>
<br>
<br>
<p>The video shows a similar comparison where predictions are made frame-by-frame. </p>
<div class="figure" style="text-align:center;">
<h4 class="">(Note: Video best seen in HD)</h4>
<iframe width="960" height="523" src="https://www.youtube.com/embed/jGdh-aq4r9A?rel=0&controls=0&showinfo=0&loop=1&playlist=jGdh-aq4r9A&modestbranding=1&autoplay=1&mute=1&start=1&vq=hd1080" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen=""></iframe>
</div>
</div>
</div>
<!-- /Row -->
</div>
<!-- /Container -->
</div>
<!-- /Sample Results -->
<!-- Energy -->
<div id="energy" class="section sm-padding">
<!-- Container -->
<div class="container">
<!-- Row -->
<div class="row">
<!-- Section header -->
<div class="section-header text-center">
<h2 class="title">Consistency Energy</h2>
</div>
<!-- /Section header -->
<div class="abstract-body col-md-12">
<p>
We quantify the amount of inconsistency in a prediction made for a query using an energy-based quantity called <strong>Consistency Energy</strong>. It is defined to be the standardized average of pairwise inconsistencies. This is also equivalent to the <strong>sample variance</strong> of the predictions. The consistency energy is an <strong>intrinsic quantity of the system</strong>, as it requires no ground truth and can be computed without any supervision. As shown below, the energy strongly correlates with supervised error and distributions shifts. This suggests the energy can be adopted as an unsuervised confidence metric.
</p>
<div class="col-sm-6">
<div class="figure">
<img width="60%" src="./img_consistency/svg/all_new.svg" alt="" class="center">
<p> Empirical analyses of consistency energy.</p>
</div>
</div>
<div class="col-sm-6">
<div class="figure"> <iframe width="480" height="249" src="https://www.youtube.com/embed/DEiSWx7-MX4?rel=0&controls=0&showinfo=0&loop=1&playlist=DEiSWx7-MX4&modestbranding=1&autoplay=1&mute=1&start=1&vq=hd1080" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen=""></iframe> <p> The video shows the energy vs two continuous domain shifts (JPEG compression, Blur). The domain shift distortions are applied on the frames. The energy turns out to be an unsupervised indicator of domain shift and generally correlates with the magnitude of shift.</p>
</div>
</div>
</div>
</div>
<!-- /Row -->
</div>
<!-- /Container -->
</div>
<!-- /Abstract -->
<!-- Transfers -->
<div id="transfers" class="section sm-padding">
<div class="container">
<div class="row">
<div class="row">
<div class="section-header text-center">
<h2 class="title">Interactive Visualizations</h2>
</div>
<p>To convey a more tangible understanding of learning with cross-task consistency constraints versus baselines, we compare the results of various configurations frame-by-frame on a youtube video. Visit the <a href="./visuals">visualizations page</a> to specify the comparison configuration of your choice and compare the performance.</p>
<!--div class="figure">
<video width="100%" height="464" playsinline autoplay center loop muted class="video-bg" id="video-bg" poster="./loading.gif">
<source src="./assets/videos/consistency3.mp4" type="video/mp4" alt="HTML5 background video">
</video>
</div-->
<br>
<button class="main-btn centering" onclick="window.location.href='./visuals';"><strong>View Interactive Visualizations</strong></button>
</div>
</div>
</div>
</div>
<!-- /Transfers -->
<!-- Pretrained -->
<div id="models" class="section sm-padding">
<!-- Container -->
<div class="container">
<div class="section-header text-center">
<h2 class="title">Live Demo on User Uploaded Images</h2>
</div>
<!-- Row -->
<div class="row">
<!-- Section header -->
<div class="col-sm-4">
<div class='col-xs-6'>
<p style='text-align: center;'>Input Image</p>
<img style='border-top: 0px' width="100%" src="assets/main_page/demo_teaser_1_input.png" alt="Demo input image 1.">
</div>
<div class='col-xs-6'>
<p style='text-align: center;'>3D Ken Burns*</p>
<!-- <p style='text-align: center;'>3D Ken Burns Effect</p> -->
<video width="100%" playsinline autoplay center loop muted poster="./loading.gif">
<source src="./assets/main_page/demo_teaser_1_vid.mp4" type="video/mp4" alt="Ken Burns effect from demo image 1">
</video>
</div>
</div>
<div class="col-sm-4">
<div class='col-xs-6'>
<p style='text-align: center;'>Input Image</p>
<img style='border-top: 0px' width="100%" src="assets/main_page/demo_teaser_3_input.png" alt="Demo input image 3.">
</div>
<div class='col-xs-6'>
<p style='text-align: center;'>3D Ken Burns*</p>
<!-- <p style='text-align: center;'>3D Ken Burns Effect</p> -->
<video width="100%" playsinline autoplay center loop muted poster="./loading.gif">
<source src="./assets/main_page/demo_teaser_3_vid.mp4" type="video/mp4" alt="Ken Burns effect from demo image 3">
</video>
</div>
</div>
<div class="col-sm-4">
<div class='col-xs-6'>
<p style='text-align: center;'>Input Image</p>
<img style='border-top: 0px' width="100%" src="assets/main_page/demo_teaser_2_input.png" alt="Demo input image 2.">
</div>
<div class='col-xs-6'>
<p style='text-align: center;'>3D Ken Burns*</p>
<!-- <p style='text-align: center;'>3D Ken Burns Effect</p> -->
<video width="100%" playsinline autoplay center loop muted poster="./loading.gif">
<source src="./assets/main_page/demo_teaser_2_vid.mp4" type="video/mp4" alt="Ken Burns effect from demo image 2">
</video>
</div>
</div>
<!-- <div class="col-sm-3"> </div> -->
</div>
<div class="row">
<!-- /Section header -->
<div class="col-sm-2 col-sm-offset-4 work">
<!-- <div class="col-lg-6 col-md-4 col-sm-4 col-xs-4 work"> -->
<img class="img-responsive" width="25%" src="./img/icon-demo1_resized.png" alt="">
<div class="overlay"></div>
<div class="work-content" style='text-align: center;'>
<h4>Try the live demo on your query image.</h4>
<div class="work-link">
<a href="./demo"><i class="fa fa-play-circle"></i></a>
</div>
</div>
<h4 style='text-align: center;'>Live Demo</h4>
</div>
<div class="col-sm-2 work">
<!-- <div class="col-lg-6 col-md-4 col-sm-4 col-xs-4 work"> -->
<img class="img-responsive" width="25%" src="./img_consistency/svg/Octicons-mark-github.svg" alt="">
<div class="overlay"></div>
<div class="work-content">
<h4>Download pretrained models.</h4>
<!-- <div class="overlay">
</div> -->
<div class="work-link">
<a href="https://github.com/EPFL-VILAB/XTConsistency"><i class="fa fa-cloud-download"></i></a>
</div>
</div>
<h4 style='text-align: center;'>Pretrained Models</h4>
</div>
</div>
<div class="col-xs-12">
<p>*Using the depth predicted by the consistency-based models and Niklaus et al. 2019 Ken Burn codeusing depth predicted by the consistency-based model</p>
</div>
<!-- </div> -->
<!-- /Row -->
</div>
<!-- /Container -->
</div>
<!-- Paper -->
<div id="paper" class="section sm-padding bg-grey">
<!-- Container -->
<div class="container">
<!-- Row -->
<div class="row">
<!-- Section header -->
<div class="col-sm-6">
<div class="section-header text-center">
<h2 class="title">Paper</h2>
</div>
<div class="row">
<div class="col-sm-6 col-xs-6">
<a href="./Cross_Task_Consistency_ARXIV2020.pdf">
<img style="width:80%" class="paper-img layered-paper-big" src=./img_consistency/paper_consistency.png>
</a>
</div>
<div class="col-sm-6 col-xs-6">
<p>Robust Learning Through Cross-Task Consistency. <br> <b>[CVPR 2020 Best Paper Award Nominee]</b>
</p>
<div class="bibtex">
<button class="outline-btn" onclick="window.location.href='https://github.com/EPFL-VILAB/XTConsistency'"><p>Github<p></button>
<button class="outline-btn" onclick="window.location.href='./Cross-Task_Consistency_supp_CVPR2020.pdf'"><p>Supplementary<p></button>
<button class="outline-btn" onclick="window.location.href='./slides'"><p>Slides<p></button>
</div>
<br><br>
<p>ARXIV 2020</p>
<div class="bibtex">
<button class="outline-btn" onclick="window.location.href='./Cross_Task_Consistency_ARXIV2020.pdf'"><p>PDF<p></button>
</div> <br>
<p>CVPR Camera Ready</p>
<div class="bibtex">
<button class="outline-btn" onclick="window.location.href='./Cross_Task_Consistency_CVPR2020.pdf'"><p>PDF<p></button>
</div>
</div>
<!--<button class="outline-btn" onclick="window.location.href='https://storage.googleapis.com/taskonomy_slides/taskonomy_slides.html'"><p>Slides<p></button>-->
</div>
</div>
<div class="col-sm-6">
<div class="section-header text-center">
<h2 class="title">CVPR 2020</h2>
</div>
<div class="cvpr-container">
<iframe class="cvpr-oral" border="0" id="" src="https://www.youtube.com/embed/dPzQusIjkQU?start=0&vq=hd1080" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen style=""></iframe>
</div>
</div>
<!-- /Row -->
</div>
<!-- /Container -->
</div>
<div id="team" class="section md-padding">
<!-- Container -->
<div class="container">
<!-- Row -->
<div class="row">
<!-- Section header -->
<div class="section-header text-center">
<h2 class="title">Team</h2>
</div>
<!-- /Section header -->
<!-- team -->
<div class="col-sm-2">
<div class="team">
<div class="team-img">
<img class="img-responsive" src="https://cs.stanford.edu/~amirz/index_files/amir18_sq.png" alt="">
<div class="overlay">
<div class="team-social">
<a href="//cs.stanford.edu/~amirz/"><i class="fa fa-globe"></i></a>
<a href="https://scholar.google.com/citations?user=RKjEFukAAAAJ&hl=en"><i class="ai ai-google-scholar big-icon"></i></a>
<a href="//github.com/amir32002"><i class="fa fa-github"></i></a>
</div>
</div>
</div>
<div class="team-content">
<h3>Amir Zamir</h3>
<span>EPFL, Stanford, UC Berkeley</span>
</div>
</div>
</div>
<!-- /team -->
<!-- team -->
<div class="col-sm-2">
<div class="team">
<div class="team-img">
<img class="img-responsive" src="./img/sasha.png" alt="">
<div class="overlay">
<div class="team-social">
<a href="//alexsax.github.io"><i class="fa fa-globe"></i></a>
<a href="//scholar.google.com/citations?user=PIq7jcUAAAAJ&hl=en"><i class="ai ai-google-scholar big-icon"></i></a>
<a href="//github.com/alexsax"><i class="fa fa-github"></i></a>
</div>
</div>
</div>
<div class="team-content">
<h3>Alexander (Sasha) Sax</h3>
<span>UC Berkeley</span>
</div>
</div>
</div>
<!-- /team -->
<!-- team -->
<div class="col-sm-2">
<div class="team">
<div class="team-img">
<img class="img-responsive" src="./img/teresa.jpg" alt="">
<div class="overlay">
<div class="team-social">
<a href="https://aserety.github.io/"><i class="fa fa-globe"></i></a>
<a href="https://scholar.google.com.tr/citations?user=jIboOyIAAAAJ&hl=en&oi=ao"><i class="ai ai-google-scholar big-icon"></i></a>
<a href="//github.com/aserety"><i class="fa fa-github"></i></a>
</div>
</div>
</div>
<div class="team-content">
<h3>Teresa Yeo</h3>
<span>EPFL</span>
</div>
</div>
</div>
<!-- /team -->
<!-- team -->
<div class="col-sm-2">
<div class="team">
<div class="team-img">
<img class="img-responsive" src="./img/oguzhan.jpg" alt="">
<div class="overlay">
<div class="team-social">
<a href="https://ofkar.github.io/"><i class="fa fa-globe"></i></a>
<a href="https://scholar.google.com.tr/citations?user=Ojtk2_MAAAAJ&hl=en"><i class="ai ai-google-scholar big-icon"></i></a>
<a href="//github.com/ofkar"><i class="fa fa-github"></i></a>
</div>
</div>
</div>
<div class="team-content">
<h3>Oğuzhan Fatih Kar</h3>
<span>EPFL</span>
</div>
</div>
</div>
<!-- /team -->
<!-- team -->
<div class="col-sm-2">
<div class="team">
<div class="team-img">
<img class="img-responsive" src="./img/nikhil.jpeg" alt="">
<div class="overlay">
<div class="team-social">
<a href="//github.com/nikcheerla"><i class="fa fa-github"></i></a>
</div>
</div>
</div>
<div class="team-content">
<h3>Nikhil Cheerla</h3>
<span>Stanford</span>
</div>
</div>
</div>
<!-- /team -->
</div>
<div class="row">
<!-- team -->
<div class="col-sm-2">
<div class="team">
<div class="team-img">
<img class="img-responsive" src="./img/rohan.jpeg" alt="">
<div class="overlay">
<div class="team-social">
<a href="//github.com/rohan-suri
"><i class="fa fa-github"></i></a>
</div>
</div>
</div>
<div class="team-content">
<h3>Rohan Suri</h3>
<span>Stanford</span>
</div>
</div>
</div>
<!-- /team -->
<!-- team -->
<div class="col-sm-2">
<div class="team">
<div class="team-img">
<img class="img-responsive" src="./img/cao.jpg" alt="">
<div class="overlay">
<div class="team-social">
</i></a>
<a href="//github.com/caozhangjie"><i class="fa fa-github"></i></a>
</div>
</div>
</div>
<div class="team-content">
<h3>Zhangjie Cao
</h3>
<span>Stanford</span>
</div>
</div>
</div>
<!-- /team -->
<!-- team -->
<div class="col-sm-2">
<div class="team">
<div class="team-img">
<img class="img-responsive" src="./img/malik.png" alt="">
<div class="overlay">
<div class="team-social">
<a href="//people.eecs.berkeley.edu/~malik/"><i class="fa fa-globe"></i></a>
<a href="//scholar.google.com/citations?user=oY9R5YQAAAAJ&hl=en"><i class="ai ai-google-scholar big-icon"></i></a>
</div>
</div>
</div>
<div class="team-content">
<h3>Jitendra Malik</h3>
<span>UC Berkeley</span>
</div>
</div>
</div>
<!-- /team -->
<!-- team -->
<div class="col-sm-2">
<div class="team">
<div class="team-img">
<img class="img-responsive" src="./img/leo2.png" alt="">
<div class="overlay">
<div class="team-social">
<a href="//geometry.stanford.edu/member/guibas/"><i class="fa fa-globe"></i></a>
<a href="https://scholar.google.com/citations?user=5JlEyTAAAAAJ&hl=en"><i class="ai ai-google-scholar big-icon"></i></a>
</div>
</div>
</div>
<div class="team-content">
<h3>Leonidas Guibas</h3>
<span>Stanford</span>
</div>
</div>
</div>
<!-- /team -->
</div>
</div>
<!-- /Container -->
</div>
<!-- /Team -->
<!-- Footer -->
<footer id="footer" class="sm-padding bg-dark">
<!-- Container -->
<div class="container">
<!-- Row -->
<div class="row">
<div class="col-md-12">
<!-- footer follow -->
<ul class="footer-follow">
<li><a href="https://github.com/EPFL-VILAB/XTConsistency"><i class="fa fa-github"></i></a></li>
</ul>
<!-- /footer follow -->
<!-- footer copyright -->
<div class="footer-copyright">
<p>Copyright © 2020. Swiss Federal Institute of Technology (EPFL). All Rights Reserved.</p>
<p>The source code of this website template is available <a href='https://github.com/EPFL-VILAB/project_website_template'>here</a> by <a href='http://vilab.epfl.ch/'>EPFL VILAB</a>.</p>
</div>
<!-- /footer copyright -->
</div>
</div>
<!-- /Row -->
</div>
<!-- /Container -->