-
Notifications
You must be signed in to change notification settings - Fork 0
/
index.html
933 lines (802 loc) · 37.6 KB
/
index.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
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
<!DOCTYPE html>
<html>
<head>
<link rel = "icon" href = "assets/title_image.png" type = "image/x-icon">
<meta charset="utf-8">
<meta name="description"
content="Cutscene: Active vision for Next Best View Planning in outdoor scenes ">
<meta name="keywords" content="cutscene,FoundVLAD, Foundation Models, Visual Place Recognition, VPR, DINOv2, DINO, SAM">
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta property="og:image" content="assets/Thumbnail.png" />
<title>Cutscene: Active vision for Next Best View Planning in outdoor scenes </title>
<script>
window.dataLayer = window.dataLayer || [];
function gtag() {
dataLayer.push(arguments);
}
gtag('js', new Date());
gtag('config', 'G-PYVRSFMDRL');
</script>
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro"
rel="stylesheet">
<link rel="stylesheet" href="./static/css/bulma.min.css">
<link rel="stylesheet" href="./static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="./static/css/bulma-slider.min.css">
<link rel="stylesheet" href="./static/css/fontawesome.all.min.css">
<link rel="stylesheet"
href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="./static/css/index.css">
<link rel="icon" href="./static/images/favicon.svg">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="./static/js/fontawesome.all.min.js"></script>
<script src="./static/js/bulma-carousel.min.js"></script>
<script src="./static/js/bulma-slider.min.js"></script>
<script src="./static/js/index.js"></script>
<script src="https://cdn.plot.ly/plotly-latest.min.js"></script>
</head>
<body>
<nav class="navbar" role="navigation" aria-label="main navigation">
<div class="navbar-brand">
<a role="button" class="navbar-burger" aria-label="menu" aria-expanded="false">
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
</a>
</div>
<div class="navbar-menu">
<div class="navbar-start" style="flex-grow: 1; justify-content: center;">
<a class="navbar-item" href="https://adityarauniyar.com/">
<span class="icon">
<i class="fas fa-home"></i>
</span>
</a>
</div>
</div>
</nav>
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-2 publication-title"><a href="https://visual-learning.cs.cmu.edu/">16824</a> Course Project: <br>
Active vision for Next Best View Planning in outdoor scenes.</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://adityarauniyar.com/">Aditya Rauniyar</a>,</span>
<span class="author-block">
<a href="https://www.linkedin.com/in/omaralama/">Omar Alama</a>,</span>
<span class="author-block">
<a href="https://yuechuanhou.com//">Yuechuan Hou</a>,</span>
<span class="author-block">
<a href="https://www.linkedin.com/in/mukul-ganwal/">Mukul Ganwal</a>,</span>
<br>
<!-- <span class="author-block">
<a href="https://theairlab.org/">Sebastian Scherer</a><sup>1</sup>,</span>
<span class="author-block">
<a href="https://robotics.iiit.ac.in/">Madhava Krishna</a><sup>2</sup>,</span>
<span class="author-block">
<a href="https://scholar.google.co.in/citations?user=oVS3HHIAAAAJ&hl=en">Sourav Garg</a><sup>4</sup></span> -->
</div>
<div class="is-size-6 publication-authors">
<span class="author-block">
<a href="https://www.cmu.edu/" style="color: rgb(179, 8, 8);">Carnegie Mellon University</a>
</span>
</div>
<!-- <div class="is-size-7 publication-authors">
<span class="author-block">* denotes equal contribution</span>
</div> -->
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<span class="link-block">
<a href="https://vcc.tech/UrbanScene3D"
class="external-link button is-normal ">
<span class="icon">
<i class="far fa-images"></i>
</span>
<span>Dataset</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- <section class="hero teaser"> -->
<!-- <div class="container is-max-desktop"> -->
<!-- <div class="hero-body"> -->
<!-- <video id="teaser" autoplay muted loop playsinline height="100%"> -->
<!-- <img source src="./data/method_viz/Splash GIF.gif" /> -->
<!-- <video id="dinov2_gardens" autoplay controls muted loop playsinline height="100%"> -->
<!-- <source src="./data/method_viz/splash_vid_compressed.mp4" -->
<!-- type="video/mp4"> -->
<!-- </video> -->
<!-- </video> -->
<!-- <h2 class="subtitle has-text-centered"> -->
<!-- <span class="coolname">cutscene</span> enables <span>universal visual place recognition (VPR) <i>anywhere</i>, <i>anytime</i> and under <i>anyview</i>.</span> -->
<!-- </h2> -->
<!-- </div> -->
<!-- </div> -->
<!-- </section> -->
<section class="hero is-small is-light">
<div class="hero-body">
<div class="container is-max-desktop">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Abstract</h2>
<div class="content has-text-justified">
<p>
In this project, we extend the exploration of autonomous robotic
tasks in the context of larger outdoor scenes. Building upon the
referenced work, we focus on planning views for these expansive
environments, addressing questions related to optimal data collection
given a set of reference images. A significant contribution of our
approach lies in the introduction of a cutscene augmentation method.
This innovative technique involves semantically dividing larger
outdoor scenes into smaller components. Our model is then trained
to predict uncertainty and RGB values for novel poses within these
segmented scenes. This cutscene augmentation method serves a dual
purpose. First, it effectively increases the size of the dataset by
a significant percentage, enhancing the efficiency of the training
process, and reducing overfitting on a scene. Second, and more importantly, it substantially increases
the accuracy of novel view predictions. By leveraging this method,
our project aims to overcome challenges associated with data
collection in large-scale outdoor scenarios, providing a valuable
contribution to the broader field of autonomous robotic tasks.
Our experiments, using both synthetic and real-world data,
demonstrate the effectiveness of our proposed uncertainty-guided
approach. The results showcase improved accuracy in scene representations
compared to baseline methods, validating the utility and generalizability
of our methodology.
</p>
</div>
</div>
</div>
</div>
<!-- Paper video. -->
<!-- <div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Video</h2>
<div class="publication-video">
<a id="overview_video"></a>
<iframe
src="./data/">
</iframe>
</div>
</div>
</div>
</div> -->
<!--/ Paper video. -->
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column">
<a id="interactive_demo"></a>
<h2 class="title is-3">Introduction and Related work</h2>
<div class="content has-text-justified">
<p>
Embodied robotic intelligence relies on active perception and
exploration, essential for various applications like robotic
manipulation, inspection, and vision-based navigation.
The autonomous collection of data plays a pivotal role in scene
understanding and subsequent tasks, with a focus on novel view
synthesis using multiple UAVs to make the process more robust [1].
However, a significant challenge lies in efficiently planning a
sequence of views for sensors, ensuring the acquisition of the
most valuable information while adhering to platform-specific
constraints and enhancing scene understanding for manipulation
tasks [2]. Addressing this challenge is crucial for enhancing the
training process, particularly in scenarios involving larger scenes.
</p>
</div>
</div>
</div>
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">Multi-UAV Data Gathering</h2>
<img id="q_image" src = "data/related_work/multi_uav_data_gathering.gif" width="500">
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">Scene Understanding for Manipulation</h2>
<img id="db_image" src = "data/related_work/nbv_manipulation.gif" width="500">
</div>
</div>
<script src = "data/trajectory_data/hawkins.js"></script>
<script src = "demo/plot.js"></script>
</div>
<p></p>
</div>
</section>
<section class="hero is-small is-light">
<div class="hero-body">
<div class="container is-max-desktop">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Contributions</h2>
<div class="content has-text-justified">
<p>
In our research, we explored 3D scene understanding,
focusing on larger outdoor environments represented by
UrbanScene3D. We integrated this environment into Blender,
generating datasets emulating camera poses and renders
akin to the DTU dataset. We conduct a viewing range sensitivity analysis
which involved various viewing ranges, creating datasets
with different parameters, including a double dome
configuration and maximum view angle changes aligned with
specified view ranges of 30, 60, and 90 degrees. Additionally,
we introduced cutscene augmentation, which augments that dataset by
extracting more subscenes from each big scene, improving the utility
of small datasets and giving the network more room to learn.
</p>
</div>
</div>
</div>
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">Training on larger outdoor scenes</h2>
<img id="db_image" src = "data/method_viz/View capture_180_renders.gif" width="500">
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">Viewing range sensitivity analysis</h2>
<img id="db_image" src = "data/method_viz/Selection_180_90_30.gif" width="500">
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">Cutscene Augmentation</h2>
<img id="db_image" src = "data/method_viz/cutscene.gif" width="500">
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column">
<a id="interactive_demo"></a>
<h2 class="title is-3">Background</h2>
<div class="content has-text-justified">
<p>
"NeRF" (Neural Radiance Fields) is a pioneering research paper
in computer vision that presents a revolutionary method for 3D
scene reconstruction and novel view synthesis. This approach
employs neural networks to model volumetric scenes from 2D images,
enabling the creation of immersive 3D environments and realistic
rendering of novel viewpoints. NeRF has had a profound impact
on various fields, including computer graphics, virtual reality,
and robotics.
<br>
"PixelNeRF" is an extension of the NeRF framework, it attempts to
generalize the NeRF network to multiple scenes by incorporating
pixel level features from input reference views.This advancement further
solidifies the potential of NeRF-based methods for tasks such as image synthesis, view
interpolation, and scene reconstruction, making them valuable
tools in computer vision and related domains.
</p>
</div>
</div>
</div>
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">NeRF</h2>
<div class="NeRF-video">
<iframe width="480" height="270" src="https://www.youtube.com/embed/JuH79E8rdKc?si=AsJpnK8WQT1vKl5q" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" allowfullscreen></iframe>
</div>
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">PixelNeRF</h2>
<div class="PixelNeRF-video">
<iframe width="480" height="270" src="https://www.youtube.com/embed/voebZx7f32g?si=K-HQnVwn3tHiTIwu" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" allowfullscreen></iframe> </div>
</div>
</div>
<script src = "data/trajectory_data/hawkins.js"></script>
<script src = "demo/plot.js"></script>
</div>
<p></p>
</div>
</section>
<section class="hero is-small is-light">
<div class="container is-max-desktop">
<div class="hero-body">
<div class="columns is-centered has-text-centered">
<div class="column">
<a id="interactive_demo"></a>
<h2 class="title is-3">Our Approach</h2>
<div class="content has-text-justified">
<p>
Our approach is based on the Next Best View Planning Using Uncertainty Estimation
in <a href = "https://arxiv.org/abs/2303.01284" style="color: blue;">Image-Based Neural Rendering (NeU-NBV) work by Jin</a> which had been built
on top of the pixelnerf framework. The NeU-NBV framework adapts pixel nerf
to next best view planning using two techniques: 1. It enhances the speed of
volumetric rendering through using an LSTM to predict the jumping distance
in ray tracing. 2. It incorporates uncertainity estimation by pushing the network
to output both a mean RGB value and a variance. The uncertainty and final RGB values
are then extracted by sampling from the log normal distribution 100 times and taking
the mean and variance of those samples.
<br/>
Furthermore, the (NeU-NBV) work adds a planning framework on top of their
adapted pixelnerf model as follows: Given two starting reference views,
they sample multiple candidate views within a limited viewing range.
They then feed all of those candidate views through the trained network which
in turn outputs both an RGB and uncertainty prediction for each pixel. The next
best view is then chosen such that it maximizes the information gain by targeting
the novel view with the highest average pixel uncertainity. The novel view is then
collected and the process is then repeated until the capture budget has been expended.
<br/>
</p>
</div>
</div>
</div>
</div>
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">Uncertainty prediction network architechture</h2>
<img id="db_image" src = "data/method_viz/network_architechture.png" width="500">
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">Uncertainity estimation loss function</h2>
<img id="db_image" src = "data/method_viz/loss_function.png" width="500">
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">NBV planning architechture</h2>
<img id="db_image" src = "data/planning/nbv_planning_arc.png" width="500">
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column">
<a id="interactive_demo"></a>
<h2 class="title is-3">Method</h2>
<div class="content has-text-justified">
<p>
Challenges arise when applying DTU pre-trained models for novel view
prediction on larger outdoor navigational scenes, which it wasn't
originally trained on. To address this, we embarked on an experimentation
journey to develop a robust training strategy. We employed the
UrbanScene3D environment, imported its .obj representation into
Blender, and scripted a custom dataset generation process aligned
with DTU's configuration. We also conducted a comprehensive
sensitivity analysis to explore various viewing angles, adjusting
scene parameters, such as varying radii for different configurations
and introducing a range of maximum view angle changes. However,
the complexity of larger scenes necessitates a more effective
training strategy, prompting us to delve into innovative techniques
like cutscene augmentation. This method involves dividing extensive
scenes into manageable chunks, enhancing scene analysis and
understanding for improved model performance.
</p>
</div>
</div>
</div>
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">Training on larger outdoor scenes</h2>
<img id="db_image" src = "data/method_viz/View capture_180_renders.gif" width="500">
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4"></h2>
<div class="content has-text-justified">
<p>
For our research, we employed a larger outdoor navigational environment
known as UrbanScene3D. To facilitate our experiments, we imported the .obj
file representing this environment into Blender. Subsequently, we set up a
custom script within Blender to generate a dataset. This dataset was designed
to emulate camera poses and renders configured similarly to the DTU dataset,
enabling us to conduct our experiments effectively and gather valuable
insights from this rich outdoor environment.
</p>
</div>
</div>
</div>
</div>
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">Setting Viewing ranges</h2>
<img id="db_image" src = "data/method_viz/Selection_180_90_30.gif" width="500">
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4"></h2>
<div class="content has-text-justified">
<p>
Utilizing the scene file within Blender, we conducted a comprehensive
sensitivity analysis to explore various viewing angles. This approach
allowed us to generate multiple datasets by manipulating the scene
parameters. Specifically, we created a double dome setup within the scene,
varying the radii for two different configurations. Moreover, we
introduced a range of maximum view angle changes for each dataset,
aligning the maximum change with the specified view range. This rigorous
analysis encompassed view ranges of 30 degrees, 60 degrees, and 90 degrees,
providing us with valuable datasets that capture a wide array of
perspectives and visual information for our research.
</p>
</div>
</div>
</div>
</div>
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">Cutscene Augmentation</h2>
<img id="db_image" src = "data/method_viz/how_to_cutscene.gif" width="500">
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4"></h2>
<div class="content has-text-justified">
<p>
Taking inspiration from simple yet effective 2D augmentation methods
such as cutout and cutmix, and stemming from the observations that
large scale outdoor scenes are scarce and such scenes are often comprised
of many smaller subscenes, we introduce cutscene augmentation.
The cutscene augmentation method increases the variation of the data
by cutting out semantically divisible subscenes out of a larger scene.
Those cutout scenes are scaled up and added as different data points
pushing the network to generalize and avoid overfitting when data
is scarce as it often is for large outdoor scenes.
</p>
</div>
</div>
</div>
</div>
<p></p>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column">
<a id="interactive_demo"></a>
<h2 class="title is-3">Experimental Setup</h2>
</div>
</div>
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="content">
<div class="content has-text-justified">
<p>
In our experimental setup, we employed a state-of-the-art approach
for training our computer vision model. We utilized a high-performance
computing environment with a 3090Ti GPU to accelerate our training
process. Our training procedure involved optimizing a range of
hyperparameters, including learning rates, and batch sizes, to achieve optimal performance. We conducted training
for a total of 200 epochs to ensure that our model converged to a
robust and accurate solution. This rigorous experimental setup allowed
us to achieve state-of-the-art results and validate the effectiveness
of our proposed approach.
</p>
</div>
</div>
</div>
<script src = "data/trajectory_data/hawkins.js"></script>
<script src = "demo/plot.js"></script>
</div>
<p></p>
</div>
</section>
<section class="hero is-small is-light">
<div class="hero-body">
<div class="columns is-centered has-text-centered">
<div class="container is-max-desktop">
<h2 class="title is-3">Viewing Range Sensitivity Analysis Results</h2>
<p>Below, we show qualitative results by fixing the reference views such that
they are at most 30 degrees apart and fixing the targetted novel view. We then
vary the model used to make the rgb and uncertainty estimation for this fixed input.
</p>
</div>
</div>
<div class="container is-max-desktop">
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">Reference Images </h2>
<img id="db_image" src = "data/viewing_range_analysis/scan0_reference_images.jpg" width="500">
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">Ground Truth </h2>
<img id="db_image" src = "data/viewing_range_analysis/scan0_ground_truth.jpg" width="167">
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
</div>
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">90deg </h2>
<img id="db_image" src = "data/viewing_range_analysis/scan0_best_90_rgb_53.jpg" width="500">
RGB
<img id="db_image" src = "data/viewing_range_analysis/scan0_best_90_uncertainty_53.jpg" width="500">
Uncertainty
<img id="db_image" src = "data/viewing_range_analysis/scan0_best_90_error_53.jpg" width="500">
Error
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">60deg </h2>
<img id="db_image" src = "data/viewing_range_analysis/scan0_best_60_rgb_53.jpg" width="500">
RGB
<img id="db_image" src = "data/viewing_range_analysis/scan0_best_60_uncertainty_53.jpg" width="500">
Uncertainty
<img id="db_image" src = "data/viewing_range_analysis/scan0_best_60_error_53.jpg" width="500">
Error
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">30deg</h2>
<img id="db_image" src = "data/viewing_range_analysis/scan0_best_30_rgb_53.jpg" width="500">
RGB
<img id="db_image" src = "data/viewing_range_analysis/scan0_best_30_uncertainty_53.jpg" width="500">
Uncertainty
<img id="db_image" src = "data/viewing_range_analysis/scan0_best_30_error_53.jpg" width="500">
Error
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
</div>
<p>
We further measure PSNR and SSIM as measures of reconstruction quality
during training. The curves show the superiority of the quality as you reduce the viewing range.
</p>
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">PSNR</h2>
<img id="db_image" src = "data/combined_psnr.png" width="500">
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">SSIM</h2>
<img id="db_image" src = "data/combined_ssim.png" width="500">
</div>
</div>
</div>
<p>
The sensitivity analysis we conducted affirms that our framework is
very sensitive to the viewing range. Having too big of a range will
affect the stability of training and the network's capability to learn, while
having too small of a range would limit the network's novel view estimation.
Note that even though the reconstruction quality reduces with a bigger viewing
range, the uncertainty still correlates well with the error which is desirable.
Thus, such models can still be utilized to estimate the next best view allbeit
less reliably as the uncertainty predictions would be highly noisy.
An interesting approach to study would be to incrementally increase the viewing
range during training to stabilize training while simultaneously pushing
the network to expand its viewing range uncertainty estimation capabilities.
We leave this idea for future work.
</p>
</div>
<br><br><br><br>
<div class="hero-body">
<div class="columns is-centered has-text-centered">
<div class="container is-max-desktop">
<h2 class="title is-3">Cutscene Augmentation Results</h2>
<p>
For qualitative results, we conducted a more challenging analysis
in which we varied the view 90 degrees in the x and y direction smoothly
from -45,-45 to +45,+45 centered on the top bird's eye view. We then fixed
the input views to be [-45,-45],[-35,-35],[+35,+35],[+45,+45].
</p>
</div>
</div>
<div class="container is-max-desktop">
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">Reference Images </h2>
<img id="db_image" src = "data/cutscene_animation/town_reference_images.jpg" width="500">
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">Ground Truth </h2>
<img id="db_image" src = "data/cutscene_animation/town_ground_truth.gif" width="167">
<!-- <img src = "data/viz_gifs/row3_col3.gif"> -->
</div>
</div>
</div>
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">60deg</h2>
<img id="db_image" src = "data/cutscene_animation/town_60deg.gif" width="500">
RGB
<img id="db_image" src = "data/cutscene_animation/town_60deg_uncertainty.gif" width="500">
Uncertainty
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">60deg + Cutscene</h2>
<img id="db_image" src = "data/cutscene_animation/town_60deg_ca.gif" width="500">
RGB
<img id="db_image" src = "data/cutscene_animation/town_60deg_ca_uncertainty.gif" width="500">
Uncertainty
</div>
</div>
</div>
<p>
We again measure PSNR and SSIM as measures of reconstruction quality
during training with cutscene augmentation.
</p>
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">PSNR</h2>
<img id="db_image" src = "data/combined_cutscene_psnr.png" width="500">
</div>
</div>
<div class="column has-text-centered">
<div class="content">
<h2 class="title is-4">SSIM</h2>
<img id="db_image" src = "data/combined_cutscene_ssim.png" width="500">
</div>
</div>
</div>
<p>
The performance improvements introduced by a very simple augmentation technique
such as cutscene are significant. Even though the model has never seen more than
60 degrees of viewing range in a single datapoint. It was able to generalize
to a broader viewing range (90 degrees) surpassing the baseline. The visualization
further serves to show how high uncertainity can hinder the NBV planning capabilities
of the model. On the right side, we can trivially observe that the cutscene model's uncertainity
is maximized at the birds eye view [0,0] whearas on the left side for the baseline,
it is unclear visually which is the next best view as it suffers from high amounts of
noise.
</p>
</div>
<br><br><br><br>
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column">
<h2 class="title is-3">Next Best View Planning Results</h2>
<p>
To test out planning performance, we took the trained model and tested
planning. We start with a random set of two views, sample from all available
ground truth views and choose the next best view based on three different policies.
We report our model as "ours", the DTU pretrained model as "DTU_baseline", and the max
view distance. We iterate until 6 views are collected and we measure
the improvement of reconstruction with each new view collected using
PSNR and SSIM as before.
</p>
</div>
</div>
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="content">
<img id="db_image" src = "data/planning/30deg_planning.png" width="1000">
</div>
<p>
As shown above, measured in reconstruction quality Our method outperformes
the geometric and DTU trained baselines successfully extending the NeU-NBV
framework to large outdoor scenes.
</p>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column">
<h2 class="title is-3">Conclusion and Future Work</h2>
<p>To conclude, in this work we have explored the application of <a href="https://arxiv.org/abs/2303.01284">NeU-NBV</a> planning to
large outdoor scenes. We conducted an evaluation on the sensitivity of the approach
to viewing angle range and showed that the method is highly sensitive. We further introduced
and evaluated a novel augmentation technique "cutscene". We demonstrated how cutscene
can significantly enhance the performance on outdoor large scenes.
In the future, we plan to further expand on the idea of cutscene possibly
automating the approach of semanticly dividing the scene into multiple subcenes
by utilizing a 2D birds eye view building/POI detector or perhaps using
geometric clustering clustering the floor vertices and exposing buildings.
Another interesting direction would be to add on to this augmentation technique by
rearranging the cutout scenes into novel configurations pushing the
variety of the scarce scene data we have even further. Furthermore,
our viewing range sensitivity analysis hints at the possiblity that the
network could benefit from a scheduled viewing range in which we increase
the range gradually as the network learns.</p>
</div>
</div>
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="content">
<p>
</p>
</div>
</div>
<div class="column has-text-centered">
<div class="content">
</div>
</div>
</div>
<p></p>
</div>
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title is-3">BibTeX</h2>
<pre><code>
@article{2023cutscene,
title={Cutscene: Active vision for Next Best View Planning in outdoor scenes},
author={Rauniyar, Aditya and Alama, Omar and Hou, Yuechuan and Ganwal, Mukul},
url = {https://adityarauniyar.com/cutscene.github.io/}
year={2023}
}
</code></pre>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="content has-text-centered">
<a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/">
<img alt="Creative Commons License" style="border-width:0" src="https://i.creativecommons.org/l/by-sa/4.0/88x31.png" />
</a>
</div>
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This website adapted from the Nerfies templates, which is licensed under a <a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/">Creative Commons Attribution-ShareAlike 4.0 International License</a>.
</p>
<p>
If you use the <a href="https://github.com/cutscene/cutscene.github.io">source code</a> of this website,
please also link back to the <a href="https://github.com/nerfies/nerfies.github.io">Nerfies source code</a> in your footer.
</p>
</div>
</div>
</div>
</div>
</footer>
</body>
</html>