-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathindex.html
461 lines (398 loc) · 17.8 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
<html><head><meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
<!--
<script src="./resources/jsapi" type="text/javascript"></script>
<script type="text/javascript" async>google.load("jquery", "1.3.2");</script>
-->
<style type="text/css">
@font-face {
font-family: 'Avenir Book';
src: url("./fonts/Avenir_Book.ttf"); /* File to be stored at your site */
}
body {
font-family: "Avenir Book", "HelveticaNeue-Light", "Helvetica Neue Light", "Helvetica Neue", Helvetica, Arial, "Lucida Grande", sans-serif;
font-weight:300;
font-size:14px;
margin-left: auto;
margin-right: auto;
width: 800px;
}
h1 {
font-weight:300;
}
h2 {
font-weight:300;
}
p {
font-weight:300;
line-height: 1.4;
}
code {
font-size: 0.8rem;
margin: 0 0.2rem;
padding: 0.5rem 0.8rem;
white-space: nowrap;
background: #efefef;
border: 1px solid #d3d3d3;
color: #000000;
border-radius: 3px;
}
pre > code {
display: block;
white-space: pre;
line-height: 1.5;
padding: 0;
margin: 0;
}
pre.prettyprint > code {
border: none;
}
.container {
display: flex;
align-items: center;
justify-content: center
}
.image {
flex-basis: 40%
}
.text {
padding-left: 20px;
padding-right: 20px;
}
.disclaimerbox {
background-color: #eee;
border: 1px solid #eeeeee;
border-radius: 10px ;
-moz-border-radius: 10px ;
-webkit-border-radius: 10px ;
padding: 20px;
}
video.header-vid {
height: 140px;
border: 1px solid black;
border-radius: 10px ;
-moz-border-radius: 10px ;
-webkit-border-radius: 10px ;
}
img.header-img {
height: 140px;
border: 1px solid black;
border-radius: 10px ;
-moz-border-radius: 10px ;
-webkit-border-radius: 10px ;
}
img.rounded {
border: 0px solid #eeeeee;
border-radius: 10px ;
-moz-border-radius: 10px ;
-webkit-border-radius: 10px ;
}
a:link,a:visited
{
color: #1367a7;
text-decoration: none;
}
a:hover {
color: #208799;
}
td.dl-link {
height: 160px;
text-align: center;
font-size: 22px;
}
.layered-paper-big { /* modified from: http://css-tricks.com/snippets/css/layered-paper/ */
box-shadow:
0px 0px 1px 1px rgba(0,0,0,0.35), /* The top layer shadow */
5px 5px 0 0px #fff, /* The second layer */
5px 5px 1px 1px rgba(0,0,0,0.35), /* The second layer shadow */
10px 10px 0 0px #fff, /* The third layer */
10px 10px 1px 1px rgba(0,0,0,0.35), /* The third layer shadow */
15px 15px 0 0px #fff, /* The fourth layer */
15px 15px 1px 1px rgba(0,0,0,0.35), /* The fourth layer shadow */
20px 20px 0 0px #fff, /* The fifth layer */
20px 20px 1px 1px rgba(0,0,0,0.35), /* The fifth layer shadow */
25px 25px 0 0px #fff, /* The fifth layer */
25px 25px 1px 1px rgba(0,0,0,0.35); /* The fifth layer shadow */
margin-left: 10px;
margin-right: 45px;
}
.layered-paper { /* modified from: http://css-tricks.com/snippets/css/layered-paper/ */
box-shadow:
0px 0px 1px 1px rgba(0,0,0,0.35), /* The top layer shadow */
5px 5px 0 0px #fff, /* The second layer */
5px 5px 1px 1px rgba(0,0,0,0.35), /* The second layer shadow */
10px 10px 0 0px #fff, /* The third layer */
10px 10px 1px 1px rgba(0,0,0,0.35); /* The third layer shadow */
margin-top: 5px;
margin-left: 10px;
margin-right: 30px;
margin-bottom: 5px;
}
.vert-cent {
position: relative;
top: 50%;
transform: translateY(-50%);
}
hr
{
border: 0;
height: 1px;
background-image: linear-gradient(to right, rgba(0, 0, 0, 0), rgba(0, 0, 0, 0.75), rgba(0, 0, 0, 0));
}
</style>
<title>DatasetDM: Synthesizing Data with Perception Annotations Using Diffusion Models</title>
</head>
<body data-new-gr-c-s-check-loaded="14.1093.0" data-gr-ext-installed="">
<br>
<center>
<span style="font-size:36px">DatasetDM: Synthesizing Data with Perception Annotations Using Diffusion Models</span><br><br><br>
</center>
<table align="center" width="800px">
<tbody><tr>
<td align="center" width="160px">
<center>
<span style="font-size:16px">Weijia Wu</a><sup>1,3</sup></span>
</center>
</td>
<td align="center" width="160px">
<center>
<span style="font-size:16px">Yuzhong Zhao<sup>2</sup></span>
</center>
</td>
<td align="center" width="160px">
<center>
<span style="font-size:16px">Hao Chen<sup>1</sup></span>
</center>
</td>
<td align="center" width="160px">
<center>
<span style="font-size:16px">Yuchao Gu<sup>3</sup></span>
</center>
</td>
<td align="center" width="160px">
<center>
<span style="font-size:16px">Rui Zhao<sup>3</sup></span>
</center>
</td>
<td align="center" width="160px">
<center>
<span style="font-size:16px">Yefei He<sup>1</sup></span>
</center>
</td></tr>
</tbody></table><br>
<table align="center" width="800px">
<tbody><tr>
<td align="center" width="160px">
<center>
<span style="font-size:16px">Hong Zhou</a><sup>1</sup></span>
</center>
</td>
<td align="center" width="160px">
<center>
<span style="font-size:16px">Mike Zheng Shou</a><sup>3 *</sup></span>
</center>
</td>
<td align="center" width="160px">
<center>
<span style="font-size:16px">Chunhua Shen</a><sup>1</sup></span>
</center>
</td></tr>
</tbody></table><br>
<table align="center" width="700px">
<tbody><tr>
<td align="center" width="50px">
<center>
<span style="font-size:16px"></span>
</center>
</td>
<td align="center" width="300px">
<center>
<span style="font-size:16px"><sup>1</sup>Zhejiang University</span>
</center>
</td>
<td align="center" width="300px">
<center>
<span style="font-size:16px"><sup>2</sup>University of Chinese Academy of Sciences</span>
</center>
</td>
</td>
<td align="center" width="300px">
<center>
<span style="font-size:16px"><sup>3</sup>Show Lab, National University of Singapore</span>
</center>
</td>
</tr></tbody></table>
<table align="center" width="700px">
<tbody><tr>
<td align="center" width="200px">
<center>
<br>
<span style="font-size:20px">Code
<a href="https://github.com/showlab/DatasetDM"> [GitHub]</a>
</span>
</center>
</td>
<td align="center" width="200px">
<center>
<br>
<span style="font-size:20px">
Paper <a href="https://arxiv.org/abs/2308.06160"> [arXiv]</a>
</span>
</center>
</td>
<td align="center" width="200px">
<center>
<br>
<span style="font-size:20px">
Cite <a href="./resources/noen"> [BibTeX]</a>
</span>
</center>
</td>
</tr></tbody>
</table>
<br><hr>
<br>
<center>
<img src="./resources/1691760099832.jpg" alt="alt text" style="width: 100%; object-fit: cover; max-width:100%;"></a>
</center>
<p style="text-align:justify; text-justify:inter-ideograph;"><left>
We introduce DatasetDM: a versatile dataset generation model featuring a perception decoder capable of producing an unlimited quantity of high-fidelity synthetic images, along with various perception annotations, including depth, segmentation, and human pose estimation
</left></p>
<br>
<center><h2> Abstract </h2> </center>
<p style="text-align:justify; text-justify:inter-ideograph;">
</p><div class="container">
<div class="text" width="400px">
<p style="text-align:justify; text-justify:inter-ideograph;">
<left>
Current deep networks are very data-hungry and benefit from training on large-scale datasets, which are often time-consuming to collect and annotate. By contrast, synthetic data can be generated infinitely using generative models such as DALL-E and diffusion models, with minimal effort and cost. In this paper, we present DatasetDM, a generic dataset generation model that can produce diverse synthetic images and the corresponding high-quality perception annotations (e.g., segmentation masks, and depth). Our method builds upon the pre-trained diffusion model and extends text-guided image synthesis to perception data generation. We show that the rich latent code of the diffusion model can be effectively decoded as accurate perception annotations using a decoder module. Training the decoder only needs less than 1% (around 100 images) manually labeled images, enabling the generation of an infinitely large annotated dataset. Then these synthetic data can be used for training various perception models for downstream tasks. To showcase the power of the proposed approach, we generate datasets with rich dense pixel-wise labels for a wide range of downstream tasks, including semantic segmentation, instance segmentation, and depth estimation. Notably, it achieves 1) state-of-the-art results on semantic segmentation and instance segmentation; 2) significantly more robust on domain generalization than using the real data alone; and state-of-the-art results in zero-shot segmentation setting; and 3) flexibility for efficient application and novel task composition (e.g., image editing).
</left></p>
</div>
</div>
<br>
<center><h2>Synthetic Data from DatasetDM (supported six task, include long-tail segmentation)</h2></center>
<center>
<img src="./resources/1691760445080.jpg" alt="alt text" style="width: 100%; object-fit: cover; max-width:100%;"></a>
</center>
<!-- <div class="container">-->
<!-- <div class="image" width="300px">-->
<!-- <center><p><img class="center" src="./resources/1691760445080.jpg" width="300px"></p></center>-->
<!-- </div>-->
<!-- <div class="text" width="250px">-->
<!-- <p> A ‘good’ mask annotation satisfy two conditions:-->
<!-- 1) class-discriminative. 2) high-resolution, precise mask.-->
<!-- -->
<!-- The average map shows the possibility for us to use for semantic segmentation,-->
<!-- where it is class-discriminative and fine-grained.-->
<!-- </p>-->
<!-- </div>-->
<!-- </div>-->
<!-- <p><img class="center" src="./resources/fig7.png" width="800px"></p>-->
<hr>
<br>
<center> <h2> How to do it (pipeline) </h2> </center>
<p><img class="left" src="./resources/1691760597135.jpg" width="800px"></p>
<p style="text-align:justify; text-justify:inter-ideograph;"><left>
DatasetDM consists of two main steps: 1) Training. Using diffusion inversion to extract the latent code from a small amount of data (around hundreds of pictures) and then train the perception decoder. 2) Text-guided data generation. A large language model such as GPT-4 is utilized to prompt infinite and diverse data generation for various downstream tasks.
</left></p>
<br>
<center> <h2> Perception Decoder </h2> </center>
<p><img class="left" src="./resources/1691760753673.jpg" width="800px"></p>
<p style="text-align:justify; text-justify:inter-ideograph;"><left>
The proposed decoder is a generalized architecture for the six supported tasks, with only minor variations required for different downstream applications, i.e., determining whether to activate certain layers.
</left></p>
<br>
<center> <h2> Prompting Text-Guided Data Generation </h2> </center>
<p><img class="left" src="./resources/1691760893860.jpg" width="800px"></p>
<p style="text-align:justify; text-justify:inter-ideograph;"><left>
We guide GPT-4 to produce diverse, and infinite prompts. For different downstream tasks and datasets, we give different guided prompts for GPT-4. For example, as for the urban scene of Cityscapes, the simple guided prompt is like `Please provide 100 language descriptions of urban driving scenes for the Cityscapes benchmark, containing a minimum of 15 words each. These descriptions will serve as a guide for Stable Diffusion in generating images.` In this approach, we collected L text prompts, which average around 100 prompts for each dataset.
</left></p>
<br>
<hr>
<center><h2>Experiment-1: Instance segmentation on COCO val2017</h2></center>
<p><b>Quantitative result for Instance segmentation on COCO val2017 </b> </p>
<p><img class="center" src="./resources/1691761060277.jpg" width="800px"></p>
<center><h2>Experiment-2: Semantic segmentation on VOC 2012</h2></center>
<p><b>Comparison with the previous methods. </b> </p>
<div class="container">
<div class="image" width="200px">
<center><p><img class="center" src="./resources/1691761142866.jpg" width="500px"></p></center>
</div>
<div class="text" width="250px">
<p> ‘R: ’ refers to the number of real data used to train.
</p>
</div>
</div>
<center><h2>Experiment-3: Semantic segmentation on Cityscapes</h2></center>
<div class="container">
<div class="image" width="400px">
<center><p><img class="center" src="./resources/1691761292905.jpg" width="500px"></p></center>
</div>
</div>
<center><h2>Experiment-4: Human Pose Estimation on COCO val2017</h2></center>
<div class="container">
<div class="image" width="850px">
<center><p><img class="center" src="./resources/1691761382370.jpg" width="800px"></p></center>
</div>
</div>
</div>
<center><h2>Experiment-5: Semantic segmentation on DeepFashion-MM</h2></center>
<div class="container">
<div class="image" width="850px">
<center><p><img class="center" src="./resources/1691761463051.jpg" width="800px"></p></center>
</div>
</div>
</div>
<center><h2>Experiment-6: Depth Estimation on NYU Depth V2 val dataset</h2></center>
<div class="container">
<div class="image" width="850px">
<center><p><img class="center" src="./resources/1691761510546.jpg" width="800px"></p></center>
</div>
</div>
</div>
<center><h2>Experiment-7: Zero-Shot Semantic Segmentation on PASCAL VOC 2012</h2></center>
<div class="container">
<div class="image" width="850px">
<center><p><img class="center" src="./resources/1691761556765.jpg" width="800px"></p></center>
</div>
</div>
</div>
<center><h2>Experiment-8: Performance for Domain Generalization between different datasets.</h2></center>
<div class="container">
<div class="image" width="850px">
<center><p><img class="center" src="./resources/1691761606825.jpg" width="800px"></p></center>
</div>
</div>
</div>
<center>
<img src="./resources/1691761637859.jpg" alt="alt text" style="width: 100%; object-fit: cover; max-width:100%;"></a>
</center>
<p style="text-align:justify; text-justify:inter-ideograph;"><left>
Examples of Human-Centric Generated Data for DatasetDM. Our method is capable
of generating high-accuracy, high-diversity, and unified perceptual annotations.
</left></p>
<br>
<center>
<img src="./resources/1691761737174.jpg" alt="alt text" style="width: 100%; object-fit: cover; max-width:100%;"></a>
</center>
<p style="text-align:justify; text-justify:inter-ideograph;"><left>
Examples of Generated Data for Urban City Scenario from DatasetDM.
</left></p>
<br>
<center>
<img src="./resources/1691761781948.jpg" alt="alt text" style="width: 100%; object-fit: cover; max-width:100%;"></a>
</center>
<p style="text-align:justify; text-justify:inter-ideograph;"><left>
Prompts of diffusion model from GPT-4. By providing some simple cues, GPT-4 can
generate a vast and diverse array of prompts.
</left></p>
<br>
<br>
<hr>
<center> <h2> Acknowledgements </h2> </center>
<p>
Based on a template by <a href="https://lipurple.github.io/">
Ziyi Li</a> and <a href="http://richzhang.github.io/">Richard Zhang</a>.
</p>
<br>
<br>
</body><grammarly-desktop-integration data-grammarly-shadow-root="true"></grammarly-desktop-integration></html>