-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
596 lines (547 loc) · 25.2 KB
/
index.html
File metadata and controls
596 lines (547 loc) · 25.2 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
<!DOCTYPE html>
<html>
<head lang="en">
<meta charset="UTF-8">
<meta http-equiv="x-ua-compatible" content="ie=edge">
<title>GLUS</title>
<meta name="description" content="">
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta property="og:image" content="">
<meta property="og:image:type" content="image/png">
<meta property="og:image:width" content="1711">
<meta property="og:image:height" content="576">
<meta property="og:type" content="website" />
<meta property="og:url" content="https://glus-video.github.io/"/>
<meta property="og:title" content="GLUS: Global-Local Reasoning Unified into A Single Large Language Model for Video Segmentation" />
<meta property="og:description" content="This paper proposes a novel framework utilizing multi-modal large language models (MLLMs) for referring video object segmentation (RefVOS).
Previous MLLM-based methods commonly struggle with the dilemma between 'Ref' and 'VOS': they either specialize in understanding a few key frames (global reasoning) or tracking objects on continuous frames (local reasoning), and rely on external VOS or frame selectors to mitigate the other end of the challenge.
However, our framework <b>GLUS</b> shows that <b>G</b>lobal and <b>L</b>ocal consistency can be <b>U</b>nified into a single video <b>S</b>egmentation MLLM: a set of sparse 'context frames' provides global information, while a stream of continuous 'query frames' conducts local object tracking.
This is further supported by jointly training the MLLM with a pre-trained VOS memory bank to simultaneously digest short-range and long-range temporal information.
To improve the information efficiency within the limited context window of MLLMs, we introduce object contrastive learning to distinguish hard false-positive objects and a self-refined framework to identify crucial frames and perform propagation.
By collectively integrating these insights, our <b>GLUS</b> delivers a simple yet effective baseline, achieving new state-of-the-art for MLLMs on the MeViS and Ref-Youtube-VOS benchmark. "/>
<meta name="twitter:card" content="summary_large_image" />
<meta name="twitter:title" content="GLUS: Global-Local Reasoning Unified into A Single Large Language Model for Video Segmentation" />
<meta name="twitter:description" content="Global-Local Reasoning Unified into A Single Large Language Model for Video Segmentation"/>
<meta name="twitter:image" content="" />
<link rel="icon" href="data:image/svg+xml,<svg xmlns=%22http://www.w3.org/2000/svg%22 viewBox=%220 0 100 100%22><text y=%22.9em%22 font-size=%2290%22></text></svg>">
<link rel="stylesheet" href="https://maxcdn.bootstrapcdn.com/bootstrap/3.3.5/css/bootstrap.min.css">
<link rel="stylesheet" href="https://maxcdn.bootstrapcdn.com/font-awesome/4.4.0/css/font-awesome.min.css">
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/codemirror/5.8.0/codemirror.min.css">
<link rel="stylesheet" href="css/app.css">
<link rel="stylesheet" href="css/fontawesome.all.min.css">
<link rel="stylesheet"
href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<script defer src="js/fontawesome.all.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/Chart.js/2.5.0/Chart.min.js"></script>
<script src="js/app.js"></script>
<script src="js/synced_video_selector.js"></script>
</head>
<body style="padding: 1%; width: 100%">
<div class="container-lg text-center" style="max-width: 1500px; margin: auto;" id="main">
<!-- <div class="container" id="main"> -->
<div class="row">
<h1 class="col-md-12 text-center">
<b>GLUS</b>: Global-Local Reasoning Unified into A Single Large Language Model for Video Segmentation</br>
</h1>
</div>
<div class="row text-center">
<div class="col-md-3">
</div>
<div class="col-md-6 text-center">
<ul class="list-inline">
<li>
<a href="https://openreview.net/profile?id=~Lang_Lin3">
Lang Lin
</a><sup></sup>*
</li>
<li>
<a href="https://openreview.net/profile?id=~Xueyang_Yu1">
Xueyang Yu
</a><sup></sup>*
</li>
<li>
<a href="https://ziqipang.github.io/">
Ziqi Pang
</a><sup></sup>*
</li>
<li>
<a href="https://yxw.cs.illinois.edu/">
Yu-Xiong Wang
</a><sup></sup>
</ul>
</div>
<div class="col-md-3">
</div>
<div class="col-md-12 text-center">
<!-- <sup>1</sup>University of Illinois Urbana-Champaign,   <sup>2</sup>Massachusetts Institute of Technology,   <sup>3</sup>Adobe Research <br>* Equal Contribution -->
<sup></sup>University of Illinois Urbana-Champaign <br>* Equal Contribution
</div>
<br>
<div class="row text-center">
<span class="link-block">
<a href="https://arxiv.org/abs/2504.07962"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
<a href="https://github.com/GLUS-video/GLUS"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Code</span>
</a>
</span>
</div>
<div class="row">
<div class="col-md-8 col-md-offset-2">
<h2>
TL;DR
</h2>
<p class="text-justify">
We propose <b>GLUS</b>, unify the distinct challenges of <b>Referring Video Object Segmentation</b>, "ref" and "vos", into a simple framework for MLLMs.
</p>
</div>
</div>
<div class="row">
<div class="col-md-8 col-md-offset-2">
<img id="teaser" width="98%" src="img/teaserfig.png">
</div>
</div>
<div class="row">
<div class="col-md-8 col-md-offset-2">
<h2>
Abstract
</h2>
<p class="text-justify">
This paper proposes a novel framework utilizing multi-modal large language models (MLLMs) for referring video object segmentation (RefVOS).
Previous MLLM-based methods commonly struggle with the dilemma between 'Ref' and 'VOS': they either specialize in understanding a few key frames (global reasoning) or tracking objects on continuous frames (local reasoning), and rely on external VOS or frame selectors to mitigate the other end of the challenge.
However, our framework <b>GLUS</b> shows that <b>G</b>lobal and <b>L</b>ocal consistency can be <b>U</b>nified into a single video <b>S</b>egmentation MLLM: a set of sparse 'context frames' provides global information, while a stream of continuous 'query frames' conducts local object tracking.
This is further supported by jointly training the MLLM with a pre-trained VOS memory bank to simultaneously digest short-range and long-range temporal information.
To improve the information efficiency within the limited context window of MLLMs, we introduce object contrastive learning to distinguish hard false-positive objects and a self-refined framework to identify crucial frames and perform propagation.
By collectively integrating these insights, our <b>GLUS</b> delivers a simple yet effective baseline, achieving new state-of-the-art for MLLMs on the MeViS and Ref-Youtube-VOS benchmark.
</p>
</div>
</div>
<br>
<div class="row">
<div class="col-md-8 col-md-offset-2">
<h2> <font color="#118ab2">GLUS</font>: Global-Local Unified Reasoning Framework for MLLMs </h2>
<p class="text-justify">
We demonstrate that unifying global and local reasoning into a single MLLM for RefVOS through the design of
context and query frames constitutes a simple yet effective baseline method for MLLM-based RefVOS models.
We further introduce plug-and-play object contrastive loss and self-refinement with key frame selectors, enabling MLLM to focus on the correct objects and most relevant frames.
</p>
<img id="glus_framework" width="98%" src="img/pipeline.png">
<p class="text-justify">
</p>
</div>
</div>
<br><br>
<div class="container">
<div class="row">
<div class="col-md-12 text-center">
<h2> Quantiative Results </h2>
</div>
</div>
<br>
<div class="row">
<div class="col-md-12 text-center">
<p class="text-justify">Our GLUS trained with only RefVOS datasets realizes competitive performance among MLLM-based approaches with datasets from diverse tasks. With expanded training datasets (ED), our GLUS achieves state-of-the-art performance across RefVOS benchmarks, especially in MeViS consisting of complex video scenarios. </p>
<!-- <img src="img/quantiative results.png" alt="Quantitative Results" class="img-fluid" style="max-width: 70%;"> -->
</div>
</div>
<style>
table {
width: 100%;
border-collapse: collapse;
margin: 20px 0;
font-family: Arial, sans-serif;
}
th, td {
border: 1px solid #ddd;
text-align: center;
padding: 8px;
}
th {
background-color: #f8f9fa;
color: #333;
}
td {
background-color: #fff;
color: #333;
}
tbody tr:nth-child(odd) td {
background-color: #f9f9f9;
}
tbody tr:hover td {
background-color: #f1f1f1;
}
td[colspan="7"] {
text-align: left;
font-weight: bold;
background-color: #e9ecef;
}
</style>
<table>
<thead>
<tr>
<th rowspan="2">Method</th>
<th colspan="3">MeViS</th>
<th colspan="3">Ref-Youtube-VOS</th>
</tr>
<tr>
<th>J&F</th>
<th>J</th>
<th>F</th>
<th>J&F</th>
<th>J</th>
<th>F</th>
</tr>
</thead>
<tbody>
<tr>
<td colspan="7">Methods without LLMs</td>
</tr>
<tr>
<td>URVOS</td>
<td>27.8</td><td>25.7</td><td>29.9</td>
<td>47.2</td><td>45.2</td><td>49.1</td>
</tr>
<tr>
<td>LBDT</td>
<td>29.3</td><td>27.8</td><td>30.8</td>
<td>49.4</td><td>48.2</td><td>50.6</td>
</tr>
<tr>
<td>MTTR</td>
<td>30.0</td><td>28.8</td><td>31.2</td>
<td>55.3</td><td>54.0</td><td>56.6</td>
</tr>
<tr>
<td>ReferFormer</td>
<td>31.0</td><td>29.8</td><td>32.2</td>
<td>62.9</td><td>61.3</td><td>64.6</td>
</tr>
<tr>
<td>OnlineRefer</td>
<td>-</td><td>-</td><td>-</td>
<td>63.5</td><td>61.6</td><td>65.5</td>
</tr>
<tr>
<td>SOC</td>
<td>-</td><td>-</td><td>-</td>
<td>67.3</td><td>65.3</td><td>69.3</td>
</tr>
<tr>
<td>TempCD</td>
<td>-</td><td>-</td><td>-</td>
<td>65.8</td><td>63.6</td><td>68.0</td>
</tr>
<tr>
<td>LoSh</td>
<td>-</td><td>-</td><td>-</td>
<td>64.2</td><td>62.5</td><td>66.0</td>
</tr>
<tr>
<td>LMPM</td>
<td>37.2</td><td>34.2</td><td>40.2</td>
<td>-</td><td>-</td><td>-</td>
</tr>
<tr>
<td>DsHmp</td>
<td>46.4</td><td>43.0</td><td>49.8</td>
<td>67.1</td><td>65.0</td><td>69.1</td>
</tr>
<tr>
<td colspan="7">Methods with LLMs</td>
</tr>
<tr>
<td>LISA-7B</td>
<td>37.2</td><td>35.1</td><td>39.4</td>
<td>53.9</td><td>53.4</td><td>54.3</td>
</tr>
<tr>
<td>LISA-13B</td>
<td>37.9</td><td>35.8</td><td>40.0</td>
<td>54.4</td><td>54.0</td><td>54.8</td>
</tr>
<tr>
<td>TrackGPT-7B</td>
<td>40.1</td><td>37.6</td><td>42.6</td>
<td>56.4</td><td>55.3</td><td>57.4</td>
</tr>
<tr>
<td>TrackGPT-13B</td>
<td>41.2</td><td>39.2</td><td>43.1</td>
<td>59.5</td><td>58.1</td><td>60.8</td>
</tr>
<tr>
<td>VideoGLAMM</td>
<td>45.2</td><td>48.1</td><td>48.2</td>
<td>-</td><td>-</td><td>-</td>
</tr>
<tr>
<td>VideoLISA-3.8B</td>
<td>44.4</td><td>41.3</td><td>47.6</td>
<td>63.7</td><td>61.7</td><td>65.7</td>
</tr>
<tr>
<td>VISA-7B</td>
<td>43.5</td><td>40.7</td><td>46.3</td>
<td>61.5</td><td>59.8</td><td>63.2</td>
</tr>
<tr>
<td>VISA-13B</td>
<td>44.5</td><td>41.8</td><td>47.1</td>
<td>63.0</td><td>61.4</td><td>64.7</td>
</tr>
<tr>
<td>ViLLa</td>
<td>-</td><td>-</td><td>-</td>
<td>66.5</td><td>64.6</td><td>68.6</td>
</tr>
<tr>
<td>GLUS (ours)</td>
<td>50.3</td><td>47.5</td><td>53.2</td>
<td>66.6</td><td>65.0</td><td>68.3</td>
</tr>
<tr>
<td><b>GLUS (ours) (ED)</b></td>
<td><b>51.3</b></td><td><b>48.5</b></td><td><b>54.2</b></td>
<td><b>67.3</b></td><td><b>65.5</b></td><td><b>69.0</b></td>
</tr>
</tbody>
</table>
</div>
<br><br>
<div class="container">
<div class="row">
<div class="col-md-12 text-center">
<h2> Qualitative Results </h2>
</div>
</div>
<br>
<div class="row">
<div class="col-md-12 text-center">
<p class="text-justify">We provide qualitative comparisons between the previous state-of-the-art (DsHmp) and our
GLUS (without extedning datasets) with the videos in MeViS. Notably, these exam-
ples illustrate three challenging aspects of RefVOS: (1) Mo-
tion Understanding: RefVOS models have to distinguish
similar objects with their motions; (2) Global Reasoning:
RefVOS models should be capable of using global reason-
ing to segment the objects presented only in a short video
clip; (3) Vision-Language Reasoning: RefVOS models
should perform vision-language unified reasoning in com-
plex scenarios. The examples demonstrate that our
GLUS effectively tackles RefVOS in challenging language-
guided segmentation cases. </p>
</div>
</div>
<div class="row">
<div class="col-md-12 text-center">
<h3> Elephant Crushed by another elephant's trunk </h3>
</div>
</div>
<div class="row">
<div class="col-md-4 text-center">
<h4> Ground Truth </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample1_gt.mp4" type="video/mp4"/>
</video>
</div>
<div class="col-md-4 text-center">
<h4> Baseline </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample1_dshmp.mp4" type="video/mp4"/>
</video>
</div>
<div class="col-md-4 text-center">
<h4> GLUS </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample1_glus.mp4" type="video/mp4"/>
</video>
</div>
</div>
<br>
<div class="row">
<div class="col-md-12 text-center">
<h3> The panda that has stayed in place with little movement </h3>
</div>
</div>
<div class="row">
<div class="col-md-4 text-center">
<h4> Ground Truth </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample2_gt.mp4" type="video/mp4"/>
</video>
</div>
<div class="col-md-4 text-center">
<h4> Baseline </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample2_dshmp.mp4" type="video/mp4"/>
</video>
</div>
<div class="col-md-4 text-center">
<h4> GLUS </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample2_glus.mp4" type="video/mp4"/>
</video>
</div>
</div>
<br>
<div class="row">
<div class="col-md-12 text-center">
<h3> The panda that took a few steps to the left </h3>
</div>
</div>
<div class="row">
<div class="col-md-4 text-center">
<h4> Ground Truth </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample3_gt.mp4" type="video/mp4"/>
</video>
</div>
<div class="col-md-4 text-center">
<h4> Baseline </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample3_dshmp.mp4" type="video/mp4"/>
</video>
</div>
<div class="col-md-4 text-center">
<h4> GLUS </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample3_glus.mp4" type="video/mp4"/>
</video>
</div>
</div>
</div>
<br>
<div class="row">
<div class="col-md-12 text-center">
<h3> White car move and turn left </h3>
</div>
</div>
<div class="row" style="width: 77%; margin: 0 auto;">
<div class="col-md-4 text-center">
<h4> Ground Truth </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample4_gt.mp4" type="video/mp4"/>
</video>
</div>
<div class="col-md-4 text-center">
<h4> Baseline </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample4_dshmp.mp4" type="video/mp4"/>
</video>
</div>
<div class="col-md-4 text-center">
<h4> GLUS </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample4_glus.mp4" type="video/mp4"/>
</video>
</div>
</div>
</div>
<br>
<div class="row">
<div class="col-md-12 text-center">
<h3> Person standing behind little girl feeding rabbit </h3>
</div>
</div>
<div class="row" style="width: 78.5%; margin: 0 auto;">
<div class="col-md-4 text-center">
<h4> Ground Truth </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample5_gt.mp4" type="video/mp4"/>
</video>
</div>
<div class="col-md-4 text-center">
<h4> Baseline </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample5_dshmp.mp4" type="video/mp4"/>
</video>
</div>
<div class="col-md-4 text-center">
<h4> GLUS </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample5_glus.mp4" type="video/mp4"/>
</video>
</div>
</div>
</div>
<br>
<div class="row">
<div class="col-md-12 text-center">
<h3> Plane moves slower </h3>
</div>
</div>
<div class="row" style="width: 78.7%; margin: 0 auto;">
<div class="col-md-4 text-center">
<h4> Ground Truth </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample6_gt.mp4" type="video/mp4"/>
</video>
</div>
<div class="col-md-4 text-center">
<h4> Baseline </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample6_dshmp.mp4" type="video/mp4"/>
</video>
</div>
<div class="col-md-4 text-center">
<h4> GLUS </h4>
<video autoplay muted loop playsinline width="100%">
<source src="videos/sample6_glus.mp4" type="video/mp4"/>
</video>
</div>
</div>
</div>
<br><br>
<br><br>
<div class="row">
<div class="col-md-8 col-md-offset-2">
<h2> Conclusions </h2>
<p class="text-justify">
We introduce a simple yet effective framework based on multimodal large language models (MLLM) for referring video object segmentation (RefVOS).
Named <b>GLUS</b>, our method establishes unified global and local reasoning in a single LLM, addressing the distinct 'Ref' and 'VOS' challenges of RefVOS.
The central design is to provide MLLM with both global (<i>context frames</i>) and local (<i>query frames</i>) contexts.
Such unified global-local reasoning is further enhanced with end-to-end optimization with VOS memory modules, which improves the consistency of <b>GLUS</b>.
Finally, <b>GLUS</b> introduces plug-and-play <i>object contrastive loss</i> and <i>pseudo-labeling</i> for key frame selection, enabling the MLLM to distinguish the correct object and frame with its limited context window.
Our <b>GLUS</b> establishes the new state of the arts on RefVOS benchmarks. We hope our baseline can inspire more systematic studies enabling MLLMs to fine-grained video understanding.
</p>
</div>
</div>
<div class="row">
<div class="col-md-8 col-md-offset-2">
<h3>
Citation
</h3>
<div class="form-group col-md-10 col-md-offset-1">
<p class="text-justify" style="background-color:#D9D9D9;">
<textarea id="bibtex" class="form-control" style="background-color:#D9D9D9;" readonly>
@inproceedings{lin2025glus,
title={GLUS: Global-Local Reasoning Unified into A Single Large Language Model for Video Segmentation},
author={Lin, Lang and Yu, Xueyang and Pang, Ziqi and Wang, Yu-Xiong},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
year={2025}
}</textarea></p>
</div>
</div>
</div>
<div class="row">
<div class="col-md-8 col-md-offset-2">
<h3>
Acknowledgements
</h3>
<p class="text-justify">
<br><br>
The website template was borrowed from <a href="http://mgharbi.com/">Michaël Gharbi</a>, <a href="https://dorverbin.github.io/refnerf">Ref-NeRF</a>, and <a href="https://reconfusion.github.io/">ReconFusion</a>.
</p>
</div>
</div>
</div>
</body>
</html>