-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathshort_slides_volumental.html
760 lines (681 loc) · 37.2 KB
/
short_slides_volumental.html
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
<!doctype html>
<html lang="en">
<head>
<link href="captionss/captionss.css" rel="stylesheet" type="text/css">
<meta charset="utf-8">
<title>Perceptual Segmentation of Visual Streams by Tracking of Objects and Parts</title>
<meta name="author" content="Jérémie Papon">
<meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no">
<link rel="stylesheet" href="reveal/css/reveal.min.css">
<link rel="stylesheet" href="pkgwtheme.css" id="theme">
<link rel="stylesheet" href="reveal/lib/css/zenburn.css">
<script>
document.write('<link rel="stylesheet" href="reveal/css/print/'+(window.location.search.match(/print-pdf/gi) ? 'pdf' : 'paper')+'.css" type="text/css" media="print">');
</script>
<!--[if lt IE 9]>
<script src="reveal/lib/js/html5shiv.js"></script>
<![endif]-->
<!--Adobe Edge Runtime-->
<meta http-equiv="X-UA-Compatible" content="IE=Edge">
<script type="text/javascript" charset="utf-8" src="edge_includes/edge.5.0.0.min.js"></script>
<style>
.edgeLoad-EDGE-3489513 { visibility:hidden; }
</style>
<!--Load HTVF animation -->
<script>
AdobeEdge.loadComposition('HTVF_anim2', 'EDGE-3489513', {
scaleToFit: "none",
centerStage: "both",
minW: "0",
maxW: "undefined",
width: "1024px",
height: "625px"
}, {dom: [ ]}, {dom: [ ]});
</script>
<!--Adobe Edge Runtime End-->
</head>
<body>
<div class="reveal">
<resetfootnotecounter></resetfootnotecounter>
<resetcitationcounter></resetcitationcounter>
<div class="slides">
<section>
<p style="text-align: center"><br><br><br>Press F11 to view this in full screen.</p>
<p style="text-align: center">Press Left/Right to advance through the presentation.</p>
<p style="text-align: center">Make sure to click on the play button for Point Clouds!</p>
<p style="text-align: center">Don't miss the vertical slides - you'll see up/down arrows on the bottom right!</p>
<p style="text-align: center">You can press the "esc" key to go to a slide overview.</p>
<br>
<p style="text-align: center"> For more information on this work, visit <a href="http://www.jeremiepapon.com/">http://www.jeremiepapon.com/ </a></p>
</section>
<section data-background="data/Buildings_Cropped_Merge.png">
<h1 style="text-shadow:3px 3px #000000">Perceptual Segmentation of Visual Streams</h1>
<h3 style="text-shadow:2px 2px #000000"> by Tracking of Objects and Parts</h3>
<footer style="font-size: 36px">
<div> <a style="color: #00FF00;text-shadow:2px 2px #000000" href="http://www.jeremiepapon.com/"><strong>Jérémie Papon</strong></a> </div>
<div style="text-shadow:2px 2px #000000">
Georg-August-Universität Göttingen <br>
Bernstein Center for Computational Neuroscience <br>
Institut für Informatik <br>
</div>
</footer>
<br>
</section>
<!-- MOTIVATION -->
<section>
<h2>How do we learn to perceive objects?</h2>
<blockquote style="font-size: 1.2em; line-height: 125%; background:">“Infants appear to perceive objects by analyzing <span class="fragment highlight-green">three-dimensional surface arrangements and motions </span>... [they] divide perceptual arrays into <span class="fragment highlight-red">units that move as connected wholes</span>, <span class="fragment highlight-blue">that move separately from one another</span>, that tend to <span class="fragment highlight-green">maintain their size and shape over motion</span>, and that tend to <span class="fragment highlight-red">act upon each other only on contact.</span>” *</blockquote>
<br>
<div class="w50" style="float: right">
<img src="data/child_colors.jpg" width="70%">
</div>
<div class="w50">
<iframe data-autoplay width="100%" height="350px" src="http://www.youtube.com/embed/7GU7-zyqAvI?loop=1&html5=1&enablejsapi=1&playlist=7GU7-zyqAvI"></iframe>
</div>
<p> How can we track these units without a-priori object knowledge? </p>
<div class='footer' >
* Spelke, Elizabeth S. "Principles of object perception." Cognitive science 14, no. 1 (1990): 29-56.
</div>
</section>
<!-- Existing Stuff -->
<section>
<section>
<h2>Parsing Video Streams -<br> Existing Methodologies</h2>
<p><strong>Semantic Event Chains</strong><cite></cite> - Represents by analyzing creation & deletion of edges in segment adjacency graph.</p>
<p>Analysis of temporal evolution of graph structure yields semantics</p>
<iframe data-autoplay height="60%" width="60%" src="http://www.youtube.com/embed/Gt5TVEcSTTE?loop=1&html5=1&enablejsapi=1&&playlist=Gt5TVEcSTTE"></iframe>
<p class="rcred"><a href="https://www.youtube.com/watch?v=Gt5TVEcSTTE">Maniac Dataset: Breakfast</a></p>
<div>This requires <span class="fragment highlight-green">a-priori knowledge</span> of objects!</div>
<br>
<div class='citation'>
<footl>
<footi>Aksoy, Eren Erdal, et al. <a href="http://www.dpi.physik.uni-goettingen.de/~eaksoye/papers/IJRR_2011.pdf">Learning the semantics of object–action relations by observation.</a> <em>The International Journal of Robotics Research</em> (2011).</footi>
</footl>
</div>
</section>
<section>
<h2>Parsing Video Streams -<br> Existing Methodologies</h2>
<p><strong>Video Object Segmentation</strong>
e.g.Abramov et al.<cite></cite>
Grundmann et al.<cite></cite></p>
<p>This parses a video into spatio-temporal volumes - “objects”</p>
<p>Core assumption means that “objects” must form <span class="fragment highlight-green">continuous spatio-temporal</span> volumes!</p>
<iframe data-autoplay height="60%" width="60%" src="http://www.youtube.com/embed/RDGxfEHHFac?loop=1&html5=1&enablejsapi=1&playlist=RDGxfEHHFac"></iframe>
<p class="rcred"><a href="http://www.videosegmentation.com/">Processed on VideoSegmentation.com</a></p>
<br>
<div class='citation'>
<footl>
<footi>Abramov et al., <a href="http://dx.doi.org/10.1109/TCSVT.2012.2199389">Real-Time Segmentation of Stereo Videos on a Portable System With a Mobile GPU,</a> <em>IEEE Transactions on Circuits and Systems for Video Technology </em>2012.</footi>
<footi>Grundmann et al., <a href="http://www.cc.gatech.edu/cpl/projects/videosegmentation/cvpr2010_videosegmentation.pdf">Efficient Hierarchical Graph Based Video Segmentation,</a><em>Computer Vision and Pattern Recognition (CVPR)</em> 2010.</footi>
</footl>
</div>
</section>
<section>
<h2>Parsing Video Streams -<br> Existing Methodologies</h2>
<iframe data-autoplay height="60%" width="60%" src="http://www.youtube.com/embed/qmM4MW5SmT4?loop=1&html5=1&enablejsapi=1&html5=1&playlist=qmM4MW5SmT4"></iframe>
<p class="rcred"><a href="http://www.videosegmentation.com/">Processed on VideoSegmentation.com</a></p>
<div>Complete failure if this assumption is violated. </div>
</section>
</section>
<!--
<section style="line-height: 135%">
<h2>Transitioning to 3D</h2>
<p>To Summarize most vexing issues </p>
<ul>
<li class="fragment">Segment into spatio-temporal volumes - <span class="fragment current-visible">cannot handle occlusions</span></li>
<li class="fragment">Divide the scene into objects <em><span class="fragment highlight-green">before</span></em> observations.</li>
<ul style="list-style-type: none"><li class="fragment">Cannot learn “object-ness” from observations</li></ul>
<li class="fragment"> Only use color - 3D geometry is not considered </li>
</ul>
<p class="fragment">To overcome some of these, we use RGB-D sensors to capture Point Clouds</p>
<div class="ctr w70">
<figure class="embed hide-smooth dark" >
<img src="data/openni_cams.jpg">
<figcaption style="font-size:0.75em">
Some OpenNI Sensors which capture RGB+D data.
</figcaption>
</figure>
</div>
<p class="rcred"><a href="http://www.pointclouds.org/">Point Cloud Library (PCL)</a></p>
</section>
-->
<section>
<h2>Overview of Methodology</h2>
<div>
<img src="data/Overall_Cropped.svg" width="100%">
</div>
<div style="position:absolute;left:785px;top:137px;">
<video src="data/cutting_supervoxels_small.mp4" width="320px" height="240px" loop class="slideautostart"></video>
</div>
</section>
<section>
<section>
<h2>Voxel Cloud Connectivity Segmentation - Supervoxels</h2>
<p> Example of Supervoxels with different seed sizes - from NYU Dataset <cite></cite> </p>
<div class="ctr w100">
<img src="data/IncreasingSeedSizePlain.svg" width="100%" >
<p class="rcred"><a href="http://www.jeremiepapon.com/cvpr-2013-supervoxels/">Papon et al. CVPR 2013</a></p>
</div>
<div class="ctr" style="width:75%">
<figure class="embed-top hide-smooth dark" >
<img src="data/VCCS_Performance.svg">
<figcaption style="font-size:1.0em">
Performance of VCCS Compared to state of the art methods
</figcaption>
</figure>
</div>
<br>
<div class="citation">
<footl>
<footi>Silberman et al., <a href="http://cs.nyu.edu/~silberman/projects/indoor_scene_seg_sup.html">Indoor Segmentation and Support Inference from RGBD Images,</a> European Conference on Computer Vision (ECCV) 2012.</footi>
</footl>
</div>
</section>
<section>
<h2>Voxel Cloud Connectivity Segmentation</h2>
<ul>
<li class="fragment">VCCS <cite></cite> is a region-growing oversegmentation technique that uses local geometry to respect object boundaries</li>
<li class="fragment">Constrained to flow across voxel connections</li>
<li class="fragment">Use color, normals, and a spatial smoothness constraint </li>
</ul>
<div style="width:50%; float:left">
<figure class="embed hide-smooth dark" >
<img src="data/test55.png"">
<figcaption style="font-size:1.0em">
Test Scene
</figcaption>
</figure>
</div>
<div style="width:50%; float:right">
<figure class="embed hide-smooth dark" >
<img src="data/supervoxel-growth.gif">
<figcaption style="font-size:1.0em">
Iterative Expansion of Supervoxels using VCCS
</figcaption>
</figure>
<div>
<p class="rcred"><a href="http://users.acin.tuwien.ac.at/arichtsfeld/?site=4">OSD Dataset</a> <a href="http://www.pointclouds.org/blog/tocs/alexandrov/index.php">Sergey Alexandrov</a></p>
<br>
<div class='citation'>
<footl>
<footi>Papon et al., <a href="http://www.jeremiepapon.com/cvpr-2013-supervoxels/">Voxel Cloud Connectivity Segmentation - Supervoxels for Point Clouds,</a> Computer Vision and Pattern Recognition (CVPR) 2013.</footi>
</footl>
</div>
</section>
<section>
<h2>Quantitative Comparison to SLIC</h2>
<div>
<figure class="embed-top hide-smooth dark" >
<img src="data/IncreasingSeedSizePlain.svg">
<figcaption style="font-size:1.0em">
VCCS Supervoxels for increasing seed size.
</figcaption>
</figure>
<p class="rcred"><a href="http://www.jeremiepapon.com/cvpr-2013-supervoxels/">Papon et al. CVPR 2013</a></p>
</div>
<div>
<figure class="embed-top hide-smooth dark" >
<img src="data/ComparisonToSLIC.svg">
<figcaption style="font-size:1.0em">
SLIC Superpixels
</figcaption>
</figure>
</div>
<br>
<div class="citation" >
<footl>
<footi>Achanta et al., <a href="http://ivrg.epfl.ch/research/superpixels">SLIC Superpixels Compared to State-of-the-art Superpixel Methods, </a> IEEE Transactions on Pattern Analysis and Machine Intelligence, 2012.</footi>
</footl>
</div>
</section>
<section>
<h2> Speed and Performance vs State of the Art </h2>
<div class="ctr" style="width:80%">
<figure class="embed-top hide-smooth dark" >
<img src="data/VCCS_Performance.svg">
<figcaption style="font-size:1.0em">
Performance of VCCS Compared to state of the art methods
</figcaption>
</figure>
</div>
<div class="ctr" style="width:80%">
<figure class="embed-top hide-smooth dark" >
<img src="data/VCCS_Speed.svg">
<figcaption style="font-size:1.0em">
Speed of VCCS Compared to state of the art methods
</figcaption>
</figure>
</div>
</section>
</section>
<section>
<h2>Supervoxels in a Point Cloud</h2>
<iframe src="http://pointclouds.org/assets/viewer/pcl_viewer.html?load=http://jpapon.github.io/data/pointclouds/cutting_demo_cloud_supervoxels.pcd&psize=3" width="1100" height="750" marginwidth="0" marginheight="0" frameborder="no" allowfullscreen="" mozallowfullscreen="" webkitallowfullscreen="">
</iframe>
</section>
<section>
<section>
<h2>LCCP Segments in a Point Cloud</h2>
<iframe src="http://pointclouds.org/assets/viewer/pcl_viewer.html?load=http://jpapon.github.io/data/pointclouds/cutting_demo_cloud_lccp.pcd&psize=3" width="1100" height="750" marginwidth="0" marginheight="0" frameborder="no" allowfullscreen="" mozallowfullscreen="" webkitallowfullscreen="">
</iframe>
</section>
<section>
<h2>Local Convexity Segmentation (LCCP)<cite></cite></h2>
<p>Use a local convexity criterion on adjacency graph edges to split graph.</p> <br>
<div class="ctr w100">
<figure class="embed reveal-smooth dark" >
<img src="data/algorithmic_flow.svg">
<figcaption style="font-size:1.0em">
Flow of segmentation: voxels to supervoxels to local convex patches.
</figcaption>
</figure>
</div>
<br>
<div class='citation'>
<footl>
<footi>Stein, S.; Schoeler, M.; <strong>Papon, J.</strong>; Wörgötter, F., <a href="http://www.jeremiepapon.com/cvpr-2014-segmentation/">Object Partitioning using Local Convexity,</a> Computer Vision and Pattern Recognition (CVPR) 2014, June 2014.</footi>
</footl>
</div>
</section>
<section>
<div class="ctr w100">
<figure class="embed hide-smooth dark" >
<img src="data/lccp_result_images1.svg">
<figcaption style="font-size:1.0em">
LCCP Example Results
</figcaption>
</figure>
</div>
</section>
<section>
<div class="ctr w100">
<figure class="embed hide-smooth dark" >
<img src="data/lccp_result_images2.svg">
<figcaption style="font-size:1.0em">
LCCP Example Results
</figcaption>
</figure>
</div>
</section>
<section>
<div class="ctr w100">
<figure class="embed-top hide-smooth dark" >
<img src="data/OSD_results.png">
<figcaption style="font-size:1.0em">
LCCP Comparison on OSD Dataset
</figcaption>
</figure>
<br>
<figure class="embed-top hide-smooth dark" >
<img src="data/NYU_results.png">
<figcaption style="font-size:1.0em">
LCCP Comparison on NYU Dataset
</figcaption>
</figure>
</div>
</section>
<section>
<h2> Can segment huge full 3D scenes efficiently. </h2>
<iframe data-autoplay height="90%" width="90%" src="http://www.youtube.com/embed/RDWK09aoWFY?loop=1&html5=1&enablejsapi=1&html5=1&playlist=RDWK09aoWFY"></iframe>
</section>
</section>
<section>
<h2>Overview of Methodology</h2>
<div>
<img src="data/Overall_Cropped.svg" width="100%">
</div>
<div style="position:absolute;left:785px;top:137px;">
<video src="data/cutting_supervoxels_small.mp4" width="320px" height="240px" loop class="slideautostart"></video>
</div>
</section>
<!-- SEQUENTIAL POINT CLOUDS -->
<section>
<section>
<h2>Sequentially Updated Octree <cite></cite> </h2>
<p> If we assume no camera motion, we can reason about why voxels “disappear” </p>
<p> Check for occlusion by ray-tracing paths from voxel to camera</p>
<iframe data-autoplay height="70%" width="70%" src="http://www.youtube.com/embed/lC-wxqZP0vE?loop=1&html5=1&enablejsapi=1&playlist=lC-wxqZP0vE"></iframe>
<p> Camera is <strong style="color:red">facing us</strong> from this perspective - notice shadows extend towards the viewer.</p>
<br><br>
<div class='citation'>
<footl>
<footi>Papon et al., <a href="http://www.jeremiepapon.com/iros-2013-video-segmentation/">Point Cloud Video Object Segmentation using a Persistent Supervoxel World-Model,</a> Intelligent Robots and Systems (IROS) 2013.</footi>
</footl>
</div>
</section>
<section>
<h2>Sequential Clouds & Occlusion Reasoning </h2>
<p> Occlusions appear as “shadows” in rendered point clouds. </p>
<p> For instance, here the lemon (which we want to keep track of) and much of the table is hidden by the bowl.</p>
<iframe data-autoplay height="70%" width="70%" src="http://www.youtube.com/embed/WZmC3Z78k3M?loop=1&html5=1&enablejsapi=1&playlist=WZmC3Z78k3M"></iframe>
<p> These blank areas limit our ability to have temporal continuity - object permanence.</p>
</section>
<section>
<h2> Pointcloud <em>without</em> Occlusion Reasoning </h2>
<iframe src="http://pointclouds.org/assets/viewer/pcl_viewer.html?load=http://jpapon.github.io/data/pointclouds/occlusion_without.pcd&psize=3" width="1100" height="750" marginwidth="0" marginheight="0" frameborder="no" allowfullscreen="" mozallowfullscreen="" webkitallowfullscreen="" style="max-width: 100%;">
</iframe>
<p>Fortunately, we can perform some <span class="fragment highlight-green">low-level reasoning</span> about occlusions.</p>
</section>
<section>
<h2> Demonstration of Occlusion Reasoning </h2>
<p> Left frame shows input data without occlusion reasoning </p>
<iframe data-autoplay height="80%" width="80%" src="http://www.youtube.com/embed/mQ1VSvwE8vE?loop=1&html5=1&enablejsapi=1&playlist=mQ1VSvwE8vE"></iframe>
<p> Right shows the same input with ray-tracing checks </p>
</section>
<section>
<h2> Pointcloud <em>with</em> Occlusion Reasoning </h2>
<iframe src="http://pointclouds.org/assets/viewer/pcl_viewer.html?load=http://jpapon.github.io/data/pointclouds/occlusion_with.pcd&psize=3" width="1100" height="750" marginwidth="0" marginheight="0" frameborder="no" allowfullscreen="" mozallowfullscreen="" webkitallowfullscreen="" style="max-width: 100%;">
</iframe>
</section>
</section>
<section>
<h2>Overview of Methodology</h2>
<div>
<img src="data/Overall_Cropped.svg" width="100%">
</div>
<div style="position:absolute;left:785px;top:137px;">
<video src="data/cutting_supervoxels_small.mp4" width="320px" height="240px" loop class="slideautostart"></video>
</div>
</section>
<!-- PARTICLE FILTERING -->
<section>
<section>
<h2>Particle filter tracking in Point Clouds</h2>
<p>Stratified Correspondence Sampling <cite></cite></p>
<figure class="embed hide-smooth dark" >
<div style="width:100%">
<img src="data/StratifiedCorrespondences.svg">
<figcaption style="font-size:1.0em">
Supervoxels are used to choose spatial strata for uniform random sampling.
</figcaption>
</div>
</figure>
<br><br>
<div class='citation'>
<footl>
<footi>Papon et al., <a href="http://www.jeremiepapon.com/wacv-2015-tracking/">Spatially Stratified Correspondence Sampling for Real-Time Point Cloud Tracking,</a> Applications of Computer Vision (WACV), 2015.</footi>
</footl>
</div>
</section>
<section>
<h2>Particle filter tracking in Point Clouds</h2>
<p> Correspondence-Based Particle Filter approach is used. </p>
<figure class="embed hide-smooth dark" >
<div style="width:100%">
<img src="data/TideModelSV.svg">
<figcaption style="font-size:1.0em">
Models used for tracking are point clouds, partitioned using supervoxels into strata for sampling.
</figcaption>
</div>
</figure>
</section>
</section>
<section>
<section>
<h2> Results in Real Application </h2>
<iframe data-autoplay height="80%" width="80%" src="http://www.youtube.com/embed/9ixDtRWhKCg?loop=1&html5=1&enablejsapi=1&playlist=9ixDtRWhKCg"></iframe>
<p class="rcred"><a href="http://www.intellact.eu/">IntellAct Project</a></p>
</section>
<section>
<h2>Results on Synthetic Benchmark</h2>
<p>
<figure class="embed-top reveal-smooth dark" >
<div style="width:100%">
<img src="data/DispErrorTide.svg">
<figcaption style="font-size:1.0em">
Plot of Displacement Error vs time per frame (ms) averaged across 50 VR Test Runs for different numbers of particles and samples per stratum.
</figcaption>
</div>
</figure>
</section>
<section>
<h2>Results on Synthetic Benchmark</h2>
<figure class="embed-top reveal-smooth dark" >
<div style="width:100%">
<img src="data/AngularErrorMilk.svg">
<figcaption style="font-size:1.0em">
Plot of Rotational Error vs time per frame (ms) averaged across 50 VR Test Runs for different numbers of particles and samples per stratum.
</figcaption>
</div>
</figure>
</section>
</section>
<section>
<h2> Results on Synthetic Benchmark <cite></cite> </h2>
<iframe data-autoplay height="80%" width="80%" src="http://www.youtube.com/embed/iEJNf53v_WU?loop=1&html5=1&enablejsapi=1&playlist=iEJNf53v_WU"></iframe>
<br><br>
<div class='citation'>
<footl>
<footi>Choi and Christensen,<a href="http://www.cc.gatech.edu/~cchoi/rgbd_obj_tracking.html" “RGB-D> Object Tracking: A Particle Filter Approach on GPU,</a> International Conference on Intelligent Robots and Systems (IROS), 2013.
</footi>
<footi>Papon et al., <a href="http://www.jeremiepapon.com/wacv-2015-tracking/">Spatially Stratified Correspondence Sampling for Real-Time Point Cloud Tracking,</a> Applications of Computer Vision (WACV), 2015.</footi>
</footl>
</div>
</section>
<!-- HIERARCHICAL SUPERVOXEL TRACKING -->
<section>
<section>
<h2> Why can't we track Supervoxels directly? </h2>
<p> Cannot track exclusively at low-level due to the “aperture problem” <cite></cite></p>
<embed src="data/flash/twosquares.swf" width="90%" height="70%"></embed>
<p class="rcred"><a href="http://web.mit.edu/persci/demos/Motion&Form/demos/download.html">MIT Perceptual Science Group</a></p>
<br>
<div class='citation'>
<footl>
<footi>McDermott, et al., <a href="http://web.mit.edu/jhm/www/Pubs/McDermott_Weiss_Adelson_2001_motion_form_beyond_junctions.pdf">Beyond junctions: Nonlocal form contraints on motion interpretation.</a> <em>Perception </em> 2001.</footi>
</footl>
</div>
</section>
<section>
<h2> Tracking Low Level Patches - <br>Why Temporal Supervoxels? </h2>
<p> Tracking low level patches would let us make temporal connections <span class="fragment highlight-green"> <em>without</em> needing to specify a-priori objects.</p>
<figure class="embed hide-smooth dark" >
<img src="data/splitting_objects.svg" >
<figcaption style="font-size:1.0em">
Splitting objects are problematic if we segment and track using a-priori models. <br> How do we label the pieces?
</figcaption>
</figure>
<p class="fragment"> We have our low level patch representation - Supervoxels. </p>
<p class="fragment"> We have en efficient tracking method. </p>
<p class="fragment" style="font-size:150%; text-align:center"> So, what's the problem? </p>
</section>
</section>
<section>
<h2>Cortical Feedback Mechanisms</h2>
<p>Humans appear to use top-down feedback mechanisms <cite></cite> </p>
<p>Feedback allows high-level areas to influence low-level vision, even receptive fields </p>
<div class="ctr" style="height:80%; width:70%">
<figure class="embed hide-smooth dark" >
<img src="data/feedback_connections.jpg" >
<figcaption style="font-size:1.0em">
Feed-forward and Feedback Mechanisms in the Human Visual Cortex
</figcaption>
</figure>
</div>
<br>
<div class='citation'>
<footl>
<footi> Gilbert and Wu Li. <a href="http://www.nature.com/nrn/journal/v14/n5/full/nrn3476.html">Top-down influences on visual processing,</a> <em>Nature Reviews Neuroscience,</em> 2013.</footi>
</footl>
</div>
</section>
<section>
<h2> Hierarchical Temporal (super)Voxel Fields (HTVF) </h2>
<div id="Stage" class="EDGE-3489513"></div>
</section>
<section>
<section>
<h2> HTVF - Cutting Video 0</h2>
<iframe data-autoplay height="100%" width="100%" src="http://www.youtube.com/embed/eckGce932GA?loop=1&html5=1&enablejsapi=1&playlist=eckGce932GA"></iframe>
</section>
<section>
<h2> HTVF - Cutting Video 1</h2>
<iframe data-autoplay height="100%" width="100%" src="http://www.youtube.com/embed/ska9IDncGqY?loop=1&html5=1&enablejsapi=1&playlist=ska9IDncGqY"></iframe>
</section>
</section>
<section>
<section>
<h2> HTVF - Occlusions - With Voxel Raytracing 3 </h2>
<iframe data-autoplay height="100%" width="100%" src="http://www.youtube.com/embed/Oq49GODldQ0?loop=1&html5=1&enablejsapi=1&playlist=Oq49GODldQ0"></iframe>
</section>
<section>
<h2> HTVF - Occlusions - Without Voxel Raytracing </h2>
<iframe data-autoplay height="100%" width="100%" src="http://www.youtube.com/embed/aYOwjk00SPs?loop=1&html5=1&enablejsapi=1&playlist=aYOwjk00SPs"></iframe>
</section>
<section>
<h2> HTVF - Occlusions - With Voxel Raytracing 0</h2>
<iframe data-autoplay height="100%" width="100%" src="http://www.youtube.com/embed/WHi0lawXgBY?loop=1&html5=1&enablejsapi=1&playlist=WHi0lawXgBY"></iframe>
</section>
<section>
<h2> HTVF - Occlusions - With Voxel Raytracing 1 </h2>
<iframe data-autoplay height="100%" width="100%" src="http://www.youtube.com/embed/Ykq0T6HQVZ0?loop=1&html5=1&enablejsapi=1&playlist=Ykq0T6HQVZ0"></iframe>
</section>
<section>
<h2> HTVF - Occlusions - With Voxel Raytracing 2 </h2>
<iframe data-autoplay height="100%" width="100%" src="http://www.youtube.com/embed/2ghgMTAGm5s?loop=1&html5=1&enablejsapi=1&playlist=2ghgMTAGm5s"></iframe>
</section>
<section>
<h2> Occlusions - Just Occlusion Filling </h2>
<iframe height="100%" width="100%" src="http://www.youtube.com/embed/lC-wxqZP0vE?loop=1&html5=1&enablejsapi=1&playlist=lC-wxqZP0vE"></iframe>
</section>
</section>
<section>
<h2>Summary</h2>
<div style="font-size: 110%; line-height: 125%">
<p>We can now create local spatio-temporal fields in point cloud video </p>
<div style="float:right">
<video src="data/cutting_supervoxels_small.mp4" width="320px" height="240px" loop class="slideautostart"></video>
</div>
<div style="width:65%">
<p class="fragment"> Importantly, these fields:</p>
<ul>
<li class="fragment">Can handle occlusions - permanence </li>
<li class="fragment">Do not need to track object instances</li>
<li class="fragment">Provides stable temporal-supervoxels that can be used for learning</li>
</ul>
</div>
</div>
<br>
<div class="fragment">
<div style="width:100%; float:left">
<p style="font-size:1.3em">Possibilities with Volumental</p>
<ul>
<li class="fragment" data-fragment-index="4">Tracking of Supervoxels</li>
<ul style="list-style-type: none">
<li class="fragment" data-fragment-index="4" >Working with deforming objects a possibility</li>
<li class="fragment" data-fragment-index="4"> Since local fields are tracked, can scan moving parts</li>
<li class="fragment" data-fragment-index="4">Future work would incorporate global sensor pose as top-down feedback to improve tracking results.</li>
</ul>
<li class="fragment" data-fragment-index="5">Real-Time Tracking</li>
<ul style="list-style-type: none">
<li class="fragment" data-fragment-index="5">Possible to do tracking on local hardware in real-time</li>
<li class="fragment" data-fragment-index="5">Can track multiple independent objects and how they interact in real time - human computer interaction?</li>
</ul>
</ul>
</div>
</div>
</section>
<section>
<h2> Questions? </h2>
<img src="data/questions.jpg" height="90%"/>
</section>
<section>
<h2>Outlook and Future Work</h2>
<div style="font-size: 110%; line-height: 145%">
<p> Many opportunities exist now that we have low-level temporal connections </p>
<ul>
<li class="fragment">Bootstrap learning - learn “objectness” from observations</li>
<li class="fragment">Higher levels in the hierarchy </li>
<ul class="fragment" style="list-style-type: none">
<li> Sensor pose - improve performance with moving camera </li>
<li> Object recognition - group parts into meaningful objects </li>
<li> Occlusion reasoning - remove self occlusion, occluded movement </li>
</ul>
<li class="fragment">Dynamic level of detail & attention</li>
<ul class="fragment" style="list-style-type: none">
<li>Less samples on large uniform surfaces </li>
<li>More samples on small irregular areas </li>
</ul>
</ul>
</div>
<br>
<div style="width:50%; float:left">
<figure class="embed hide-smooth dark" >
<img src="data/icub_blocks.jpg">
<figcaption style="font-size:1.0em">
Bootstrapping Visual Understanding...
</figcaption>
</figure>
<p class="rcred"><a href="http://www.icub.org/">iCub</a></p>
</div>
<div style="width:50%; float:right">
<br>
<figure class="fragment embed hide-smooth dark" data-fragment-index="1" >
<img src="data/icub_block_evil.jpg">
<figcaption style="font-size:1.0em">
So they never lose track of you.
</figcaption>
</figure>
<p class="rcred"><a href="http://www.icub.org/">iCub</a></p>
</div>
</section>
<section>
<section>
<h2> HTVF - Camera Pan 0 </h2>
<iframe data-autoplay height="100%" width="100%" src="http://www.youtube.com/embed/0D9dOgv9COw?loop=1&html5=1&enablejsapi=1&playlist=0D9dOgv9COw"></iframe>
</section>
<section>
<h2> HTVF - Camera Pan 2 - LCCP Overlay</h2>
<iframe data-autoplay height="100%" width="100%" src="http://www.youtube.com/embed/97MJs3MREPo?loop=1&html5=1&enablejsapi=1&playlist=97MJs3MREPo"></iframe>
</section>
<section>
<h2> HTVF - Camera Pan 1</h2>
<iframe data-autoplay height="100%" width="100%" src="http://www.youtube.com/embed/uYKIWuRshkQ?loop=1&html5=1&enablejsapi=1&playlist=uYKIWuRshkQ"></iframe>
</section>
</section>
<!-- End of slides. -->
<script src="reveal/lib/js/head.min.js"></script>
<script src="reveal/js/reveal.min.js"></script>
<script src="pdfjs/compatibility.js"></script>
<script src="pdfjs/pdf.js"></script>
<script>
Reveal.initialize({
width: 1200,
height: 800,
controls: true,
progress: true,
history: true,
center: true,
keyboard: true,
overview: true,
theme: Reveal.getQueryHash().theme, // available themes are in /css/theme
transition: Reveal.getQueryHash().transition || 'default', // none/fade/slide/convex/concave/zoom
transitionSpeed: 'default', // default/fast/slow
math: {
mathjax: 'mathjax/MathJax.js',
config: 'TeX-AMS_HTML-full',
},
dependencies: [
{ src: 'reveal/lib/js/classList.js',
condition: function() { return !document.body.classList; }},
{ src: 'reveal/plugin/markdown/marked.js',
condition: function() { return !!document.querySelector ('[data-markdown]'); }},
{ src: 'reveal/plugin/markdown/markdown.js',
condition: function() { return !!document.querySelector ('[data-markdown]'); }},
{ src: 'reveal/plugin/highlight/highlight.js', async: true,
callback: function() { hljs.initHighlightingOnLoad (); }},
{ src: 'reveal/plugin/notes/notes.js', async: true,
condition: function() { return !!document.body.classList; }},
{ src: 'mymath.js', async: true },
{ src: 'pdfimgs.js', async: true },
{ src: 'slideautostart.js', async: true },
],
});
// Only load SDO Data page if we go to that slide
Reveal.addEventListener ('sdodataslide', function () {
console.log('sdodataslide');
document.querySelector ('#sdoiframe').src = 'http://sdo.gsfc.nasa.gov/data/'
});
</script>
</body>
</html>