-
Notifications
You must be signed in to change notification settings - Fork 0
/
index.html
475 lines (416 loc) · 16 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
<!DOCTYPE html>
<html>
<head>
<!-- Standard Meta -->
<meta charset="utf-8" />
<meta http-equiv="X-UA-Compatible" content="IE=edge,chrome=1" />
<meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=1.0">
<!-- Site Properties -->
<title>Airbert - ICCV 2021</title>
<!-- SEO -->
<meta property="og:title" content="Airbert: In-domain Pretraining for Vision-and-Language Navigation" />
<meta property="og:type" content="article" />
<meta property="og:description" content="SOTA in multiple VLN tasks by pre-training on Airbnb" />
<meta property="og:image" content="https://airbert-vln.github.io/assets/img/teaser.jpeg" />
<meta property="og:url" content="https://airbert-vln.github.io/" />
<!-- Twitter Card data -->
<meta name="twitter:card" content="summary" />
<meta name="twitter:title" content="Airbert: In-domain Pretraining for Vision-and-Language Navigation" />
<meta name="twitter:description" content="SOTA in multiple VLN tasks by pre-training on Airbnb" />
<meta name="twitter:image" content="https://airbert-vln.github.io/assets/img/teaser_square.jpeg" />
<!-- You MUST include jQuery before Fomantic -->
<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/jquery.min.js"></script>
<link rel="stylesheet" type="text/css" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/semantic.min.css">
<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/semantic.min.js"></script>
<script src="https://polyfill.io/v3/polyfill.min.js?features=es6"></script>
<script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js"></script>
<style type="text/css">
.hidden.menu {
display: none;
}
.masthead.segment {
min-height: 700px;
padding: 1em 0em;
}
.masthead .logo.item img {
margin-right: 1em;
}
.masthead .ui.menu .ui.button {
margin-left: 0.5em;
}
.masthead h1.ui.header {
margin-top: 3em;
margin-bottom: 0em;
font-size: 4em;
font-weight: normal;
}
.masthead h2 {
font-size: 1.7em;
font-weight: normal;
}
/
.ui.vertical.stripe {
padding: 8em 0em;
}
.ui.vertical.stripe h3 {
font-size: 2em;
}
.ui.vertical.stripe .button + h3,
.ui.vertical.stripe p + h3 {
margin-top: 3em;
}
.ui.vertical.stripe .floated.image {
clear: both;
}
.ui.vertical.stripe p {
font-size: 1.33em;
}
.ui.vertical.stripe .horizontal.divider {
margin: 3em 0em;
}
.quote.stripe.segment {
padding: 0em;
}
.quote.stripe.segment .grid .column {
padding-top: 5em;
padding-bottom: 5em;
}
.footer.segment {
padding: 5em 0em;
}
.secondary.pointing.menu .toc.item {
display: none;
}
@media only screen and (max-width: 700px) {
.ui.fixed.menu {
display: none !important;
}
.secondary.pointing.menu .item,
.secondary.pointing.menu .menu {
display: none;
}
.secondary.pointing.menu .toc.item {
display: block;
}
.masthead.segment {
min-height: 350px;
}
.masthead h1.ui.header {
font-size: 2em;
margin-top: 1.5em;
}
.masthead h2 {
margin-top: 0.5em;
font-size: 1.5em;
}
}
p {
text-align: justify;
font-size: 12pt;
}
.masthead {
background-image: url('/assets/img/bg3.jpg') !important;
background-size: cover !important;
}
.masthead.segment {
min-height: 300px;
}
.masthead h1.ui.header {
margin-top: 0em;
}
.masthead .ui.tex a {
margin-bottom: 40px;
}
.masthead a {
color: #EEE;
}
.ui.vertical.stripe.segment {
padding: 5em 0em;
}
</style>
<script>
$(document)
.ready(function() {
// fix menu when passed
$('.masthead')
.visibility({
once: false,
onBottomPassed: function() {
$('.fixed.menu').transition('fade in');
},
onBottomPassedReverse: function() {
$('.fixed.menu').transition('fade out');
}
})
;
// create sidebar and attach to menu open
$('.ui.sidebar')
.sidebar('attach events', '.toc.item')
;
})
;
</script>
</head>
<body>
<!-- Following Menu -->
<div class="ui large top fixed hidden menu">
<div class="ui container">
<a href="index.html" class="active item">
<i class="home icon"></i>Home
</a>
<a href="demo.html" class="item">
<i class="robot icon"></i>Demo
</a>
<a href="paper.html" class="item">
<i class="book icon"></i>Paper
</a>
<a href="https://arxiv.org/abs/2108.09105" class="item">
<i class="glasses icon"></i>arXiv
</a>
<a href="bibtex.txt" class="item">
<i class="quote right icon"></i>BibTeX
</a>
<a href="https://github.com/airbert-vln" class="item">
<i class="github icon"></i>GitHub
</a>
</div>
</div>
<!-- Sidebar Menu -->
<div class="ui vertical inverted sidebar menu">
<a href="index.html" class="active item">
<i class="home icon"></i>Home
</a>
<a href="demo.html" class="item">
<i class="robot icon"></i>Demo
</a>
<a href="paper.html" class="item">
<i class="book icon"></i>Paper
</a>
<a href="https://arxiv.org/abs/2108.09105" class="item">
<i class="glasses icon"></i>arXiv
</a>
<a href="bibtex.txt" class="item">
<i class="quote right icon"></i>BibTeX
</a>
<a href="https://github.com/airbert-vln" class="item">
<i class="github icon"></i>GitHub
</a>
<a href="https://www.youtube.com/watch?v=veND1vIkdm" class="item">
<i class="youtube icon"></i>Video
</a>
</div>
<!-- Page Contents -->
<div class="pusher">
<div class="ui inverted vertical masthead center aligned segment">
<div class="ui large secondary inverted pointing menu">
<div class="ui container">
<a class="toc item">
<i class="sidebar icon"></i>
</a>
<a href="index.html" class="active item">
<i class="home icon"></i>Home
</a>
<a href="demo.html" class="item">
<i class="robot icon"></i>Demo
</a>
<a href="paper.html" class="item">
<i class="book icon"></i>Paper
</a>
<a href="https://arxiv.org/abs/2108.09105" class="item">
<i class="glasses icon"></i>arXiv
</a>
<a href="bibtex.txt" class="item">
<i class="quote right icon"></i>BibTeX
</a>
<a href="https://github.com/airbert-vln" class="item">
<i class="github icon"></i>GitHub
</a>
<a href="https://www.youtube.com/watch?v=veND1vIkdm" class="item">
<i class="youtube icon"></i>Video
</a>
</div>
</div>
<div class="ui text container">
<h1 class="ui inverted header">
Airbert
</h1>
<h2>
In-domain Pretraining for Vision-and-Language Navigation
</h2>
<h4>
<a href="https://www.linkedin.com/in/pierre-louis-guhur-51130495/">Pierre-Louis Guhur</a> <sup>1, 2</sup>,
<a href="https://makarandtapaswi.github.io/">Makarand Tapaswi</a> <sup>3 </sup> ,
<a href="https://cshizhe.github.io/">Shizhe Chen</a> <sup>1</sup>,
<a href="https://www.di.ens.fr/~laptev/">Ivan Laptev <sup>1, 2</sup></a>,
<a href="https://www.di.ens.fr/willow/people_webpages/cordelia/">Cordelia Schmid</a> <sup>1, 2</sup>
</h4>
<h4>
<sup>1</sup> <a href="https://www.inria.fr"> Inria Paris</a>,
<sup>2</sup> <a href="https://www.di.ens.fr/">ENS, CNRS, PSL Research University</a>,
<sup>2</sup> <a href=https://www.iiit.ac.in">IIIT Hyderabad</a>
</h4>
</div>
</div>
<div class="ui vertical stripe segment" id="motivation">
<div class="ui middle aligned stackable grid container">
<div class="row">
<div class="eight wide column">
<h3 class="ui header">
<div class="ui red horizontal big label">
Problem
</div>
How to follow instructions in environments with new objects?<em data-emoji=':christmas_tree:'></em>
</h3>
<p>
VLN tasks are evaluated on unseen environments at test time.
These environments contain new objects.
How can an agent follow an instruction referring to a Christmas tree when the latter has never been observed in the language or visual corpus?
</p>
</div>
<div class="eight wide right floated column">
<img src="assets/img/christmas_tree.svg" class="ui massive image">
</div>
</div>
<div class="row">
<div class="eight wide column">
<h3 class="ui header">
<div class="ui green horizontal big label">
Solution
</div>
Build a large-scale dataset with navigation instructions from <a href="https://github.com/airbert-vln/bnb-dataset/">BnB</a> listings
<em data-emoji=":earth_americas:" ></em>
</h3>
<p>We build a large-scale, visually diverse, and in-domain dataset by creating path-instruction pairs close to a VLN-like setup and show the benefits of self-supervised pretraining.</p>
</div>
<div class="eight wide right floated column">
<img src="assets/img/bnb-dataset.svg" class="ui massive image">
</div>
</div>
<div class="row center aligned">
<div class="fourteen wide column">
<a href="demo.html" class="ui primary large labeled icon button"><i class="align robot icon"></i> Demo</a>
<a href="paper.html" class="ui primary large labeled icon button"><i class="align book icon"></i>Paper</a>
<button class="ui primary large labeled icon button"><i class="align glasses icon"></i>arXiv</button>
<button class="ui primary large labeled icon button"><i class="align quote right icon"></i>BibTeX</button>
<a href="https://github.com/airbert-vln" class="ui primary large labeled icon button">
<i class="github icon"></i>GitHub
</a>
<a href="https://www.youtube.com/watch?v=veND1vIkdm" class="item">
<i class="youtube icon"></i>Video
</a>
</div>
</div>
</div>
</div>
<div class="ui vertical stripe segment" >
<div class="ui middle aligned stackable grid container">
<h1 class="ui header">Building path-instruction pairs from BnB dataset</h1>
<p>Though navigation instructions are rarely found on the Internet, image-caption pairs from home environments are abundant in online marketplaces (<em>e.g</em>. <a href="airbnb.com">Airbnb</a>), which include images and descriptions of rental listings.</p>
<p>We collect BnB, a new large-scale dataset with 1.4M indoor images and 0.7M captions. First, we show that in-domain image-caption pairs bring additional benefits for downstream VLN tasks when applied with generic web data <span class="citation" data-cites="majumdar2020vlnbert">[17]</span>. In order to further reduce the domain gap between the BnB pretraining and the VLN task, we present an approach to transform static image-caption pairs into visual paths and navigation-like instructions, leading to large additional performance gains.
<div class="ui center aligned raised very padded text container segment">
<img class="ui fluid image" src="/assets/img/dataset.svg" />
</div>
</div>
</div>
<div class="ui vertical stripe segment" >
<div class="ui middle aligned stackable grid container">
<h1 class="ui header">Airbert: A Pretrained VLN Model</h1>
<img class="ui fluid image" src="/assets/img/pretraining.svg" />
<p>Our pretrained model, Airbert, is a generic transformer backbone that can be readily integrated in both discriminative VLN tasks such as path-instruction compatibility prediction <span class="citation" data-cites="majumdar2020vlnbert">[17]</span> and generative VLN tasks <span class="citation" data-cites="hong2021recurrentvln">[18]</span> in R2R navigation <span class="citation" data-cites="anderson2018r2r">[2]</span> and REVERIE remote referring expression <span class="citation" data-cites="qi2020reverie">[19]</span>. We achieve state-of-the-art performance on these VLN tasks with our pretrained model. Beyond the standard evaluation, our in-domain pretraining opens an exciting new direction of <em>one/few-shot VLN</em> where the agent is trained on examples only from one/few environment(s) and expected to generalize to other unseen environments.</p>
<p> We also propose a shuffling loss that improves the model’s temporal reasoning abilities by learning a temporal alignment between a path and the corresponding instruction.</p>
</div>
</div>
<div class="ui vertical stripe segment" >
<div class="ui middle aligned stackable grid container">
<h1 class="ui header">Results</h1>
<p>
We evaluate Airbert on the test set by submitting our best method to the R2R leaderboard. As seen on the following Table, our method achieves the highest success rate at 77% and is ranked first as of the submission deadline.
</p>
<table class="ui small striped table">
<tbody>
<thead class="even">
<th style="text-align: left;"></th>
<th style="text-align: center;">PL</th>
<th style="text-align: center;">NE</th>
<th style="text-align: center;">SPL</th>
<th style="text-align: center;">OSR</th>
<th style="text-align: center;">SR</th>
</thead>
<tr class="odd">
<td style="text-align: left;">Speaker-Follower <span class="citation" data-cites="fried2018speaker">[27]</span></td>
<td style="text-align: center;">1,257</td>
<td style="text-align: center;">4.87</td>
<td style="text-align: center;">0.01</td>
<td style="text-align: center;">96</td>
<td style="text-align: center;">53</td>
</tr>
<tr class="even">
<td style="text-align: left;">PreSS <span class="citation" data-cites="li2019press">[16]</span></td>
<td style="text-align: center;">10.5</td>
<td style="text-align: center;">24.5</td>
<td style="text-align: center;">0.63</td>
<td style="text-align: center;">57</td>
<td style="text-align: center;">53</td>
</tr>
<tr class="odd">
<td style="text-align: left;">PREVALENT <span class="citation" data-cites="hao2020prevalent">[14]</span></td>
<td style="text-align: center;">10.21</td>
<td style="text-align: center;">4.52</td>
<td style="text-align: center;">0.56</td>
<td style="text-align: center;">64</td>
<td style="text-align: center;">59</td>
</tr>
<tr class="even">
<td style="text-align: left;">Self-Monitoring <span class="citation" data-cites="ma2019self">[28]</span></td>
<td style="text-align: center;">373</td>
<td style="text-align: center;">4.48</td>
<td style="text-align: center;">0.02</td>
<td style="text-align: center;">97</td>
<td style="text-align: center;">61</td>
</tr>
<tr class="odd">
<td style="text-align: left;">Reinforced CM <span class="citation" data-cites="wang2019reinforced">[31]</span></td>
<td style="text-align: center;">358</td>
<td style="text-align: center;">4.03</td>
<td style="text-align: center;">0.02</td>
<td style="text-align: center;">96</td>
<td style="text-align: center;">63</td>
</tr>
<tr class="even">
<td style="text-align: left;">EnvDrop <span class="citation" data-cites="anderson2018r2r">[2]</span></td>
<td style="text-align: center;">687</td>
<td style="text-align: center;">3.26</td>
<td style="text-align: center;">0.01</td>
<td style="text-align: center;">99</td>
<td style="text-align: center;">69</td>
</tr>
<tr class="odd">
<td style="text-align: left;">AuxRN <span class="citation" data-cites="zhu2020auxrn">[51]</span></td>
<td style="text-align: center;">41</td>
<td style="text-align: center;">3.24</td>
<td style="text-align: center;">0.21</td>
<td style="text-align: center;">81</td>
<td style="text-align: center;">71</td>
</tr>
<tr class="even">
<td style="text-align: left;">VLN-BERT <span class="citation" data-cites="majumdar2020vlnbert">[17]</span></td>
<td style="text-align: center;">687</td>
<td style="text-align: center;">3.09</td>
<td style="text-align: center;">0.01</td>
<td style="text-align: center;">99</td>
<td style="text-align: center;">73</td>
</tr>
<tr style="font-weight: bold;" class="odd">
<td style="text-align: left">Airbert (ours)</td>
<td style="text-align: center;">687</td>
<td style="text-align: center;">2.69</td>
<td style="text-align: center;">0.01</td>
<td style="text-align: center;">99</td>
<td style="text-align: center;">77</td>
</tr>
</tbody>
</table>
</div>
</div>
</body>
</html>