-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathindex.html
561 lines (500 loc) · 27 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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>DitHub: Incremental Open-Vocabulary Object Detection</title>
<link href="https://cdnjs.cloudflare.com/ajax/libs/bootstrap/5.3.2/css/bootstrap.min.css" rel="stylesheet">
<link href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.5.1/css/all.min.css" rel="stylesheet">
<link rel="apple-touch-icon" sizes="180x180" href="icon/apple-touch-icon.png">
<link rel="icon" type="image/png" sizes="32x32" href="icon/favicon-32x32.png">
<link rel="icon" type="image/png" sizes="16x16" href="icon/favicon-16x16.png">
<link rel="manifest" href="icon/site.webmanifest">
<link href="https://fonts.googleapis.com/css2?family=Inter:wght@300;400;500;700&family=Montserrat:wght@300;400;500;600;700&display=swap" rel="stylesheet">
<style>
:root {
--primary-color: #505050;
--secondary-color: #2C3E50;
--accent-color: #707070;
--light-gray: #f8f9fa;
--dark-gray: #495057;
--text-color: #333;
--text-secondary: #6c757d;
}
body {
font-family: 'Inter', sans-serif;
color: var(--text-color);
line-height: 1.6;
background-color: #fff;
}
h1, h2, h3, h4, h5, h6 {
font-family: 'Montserrat', sans-serif;
font-weight: 600;
}
.btn-paper {
margin: 0 10px;
border-radius: 30px;
padding: 10px 20px;
font-weight: 500;
transition: all 0.3s ease;
text-transform: uppercase;
letter-spacing: 1px;
font-size: 0.85rem;
}
.btn-paper:hover {
transform: translateY(-3px);
box-shadow: 0 10px 20px rgba(0, 0, 0, 0.1);
}
.disabled {
pointer-events: none;
opacity: 0.6;
}
.hero-section {
padding: 6rem 0;
background: linear-gradient(135deg, #f8f9fa 0%, #e9ecef 100%);
color: #272727;
position: relative;
overflow: hidden;
}
.hero-section::before {
content: "";
position: absolute;
top: 0;
left: 0;
width: 100%;
height: 100%;
background: url('data:image/svg+xml;base64,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');
opacity: 0.8;
}
.hero-section h1,
.hero-section h2,
.hero-section p,
.hero-section .authors {
color: #272727;
}
.hero-section h2 {
opacity: 1;
}
.authors {
color: #272727;
font-size: 1.1rem;
font-weight: 300;
}
.btn-paper {
margin: 0 10px;
border-radius: 30px;
padding: 10px 20px;
font-weight: 500;
transition: all 0.3s ease;
text-transform: uppercase;
letter-spacing: 1px;
font-size: 0.85rem;
}
.btn-light {
color: #FFFFFF;
background-color: #b31b1b;
border-color: #b31b1b;
}
.btn-light:hover {
background-color: #b31b1b;
border-color: #b31b1b;
color: #FFFFFF;
transform: translateY(-3px);
box-shadow: 0 10px 20px rgba(0, 0, 0, 0.1);
}
.btn-outline-light {
color: #FFFFFF;
border-color: #FFFFFF;
background-color: transparent;
}
.btn-outline-light:hover {
background-color: rgba(255, 255, 255, 0.2);
color: #FFFFFF;
transform: translateY(-3px);
box-shadow: 0 10px 20px rgba(0, 0, 0, 0.1);
}
.table-results {
font-size: 0.80rem;
border-radius: 10px;
overflow: hidden;
box-shadow: 0 5px 15px rgba(0, 0, 0, 0.05);
width: 100%;
}
.table-results th {
background-color: var(--primary-color);
color: white;
font-weight: 500;
padding: 12px 8px;
white-space: nowrap;
}
.table-results td {
padding: 10px 8px;
white-space: nowrap;
}
.best-result {
font-weight: bold;
color: #198754;
}
.section-title {
position: relative;
margin-bottom: 2rem;
display: inline-block;
}
.section-title::after {
content: "";
position: absolute;
left: 0;
bottom: -10px;
width: 50px;
height: 4px;
background-color: var(--primary-color);
border-radius: 2px;
}
.feature-card {
border-radius: 15px;
border: none;
overflow: hidden;
transition: transform 0.3s ease, box-shadow 0.3s ease;
height: 100%;
}
.feature-card:hover {
transform: translateY(-5px);
box-shadow: 0 15px 30px rgba(0, 0, 0, 0.1);
}
.feature-card .card-title {
font-weight: 600;
color: var(--primary-color);
}
.feature-card .card-icon {
font-size: 2rem;
margin-bottom: 1rem;
color: var(--primary-color);
}
.content-section {
padding: 5rem 0;
}
.light-section {
background-color: var(--light-gray);
}
.lead {
font-size: 1.15rem;
font-weight: 400;
color: var(--text-secondary);
line-height: 1.8;
}
.highlight-text {
color: var(--primary-color);
font-weight: 500;
}
.model-diagram {
border-radius: 15px;
box-shadow: 0 15px 30px rgba(0, 0, 0, 0.1);
transition: transform 0.3s ease;
}
.model-diagram:hover {
transform: scale(1.02);
}
.blob-background {
background: linear-gradient(135deg, rgba(200, 200, 200, 0.1) 0%, rgba(220, 220, 220, 0.1) 100%);
}
@keyframes blob {
0% { border-radius: 43% 57% 70% 30% / 46% 40% 60% 54%; }
25% { border-radius: 54% 46% 38% 62% / 49% 70% 30% 51%; }
50% { border-radius: 30% 70% 50% 50% / 30% 30% 70% 70%; }
75% { border-radius: 65% 35% 65% 35% / 40% 60% 40% 60%; }
100% { border-radius: 43% 57% 70% 30% / 46% 40% 60% 54%; }
}
</style>
</head>
<body>
<!-- Hero Section -->
<div class="hero-section position-relative">
<div class="container position-relative z-index-1">
<div class="row justify-content-center text-center">
<div class="col-md-10">
<h1 class="display-3 mb-3 d-flex align-items-center justify-content-center fw-bold" style="font-family: 'Montserrat', sans-serif;">
<img src="logo.png" class="me-3" style="height: 80px;">
<span>DitHub</span>
</h1>
<h2 class="h4 mb-4 text-black opacity-75 fw-light">A Modular Framework for Incremental Open-Vocabulary Object Detection</h2>
<p class="authors mb-4">
Chiara Cappellino, Gianluca Mancusi, Matteo Mosconi, Angelo Porrello, Simone Calderara, Rita Cucchiara
</p>
<div class="buttons d-flex justify-content-center mt-3">
<a href="http://arxiv.org/abs/2503.09271" class="btn btn-light btn-paper shadow"><i class="fas fa-archive me-2"></i>arXiv Paper</a>
<button class="btn btn-dark btn-paper disabled ms-2">
<i class="fab fa-github me-2"></i>Code (coming soon)
</button>
</div>
</div>
</div>
</div>
<div class="blob-background" style="top: -400px; right: -400px; opacity: 0.5;"></div>
<div class="blob-background" style="bottom: -400px; left: -400px; opacity: 0.5;"></div>
</div>
<!-- Main Content -->
<div class="content-section">
<div class="container">
<!-- Abstract -->
<div class="row justify-content-center mb-5">
<div class="col-md-10">
<h3 class="section-title">Abstract</h3>
<p class="lead">
Open-Vocabulary object detectors inherently generalize to an unrestricted set of categories, enabling recognition through simple textual prompting. However, adapting these models
to rare classes or reinforcing their abilities on specialized domains remains essential. While recent methods rely on monolithic adaptation strategies with a single set of weights,
we embrace <span class="highlight-text">modular deep learning</span>. We introduce <span class="highlight-text">DitHub</span>, a framework designed to build and maintain a library
of efficient adaptation modules. Inspired by Version Control Systems, DitHub manages expert modules as branches that can be fetched and merged as needed. This modular approach
allows us to conduct an in-depth exploration of the compositional properties of adaptation modules, marking the first such study in Object Detection. Our method achieves
state-of-the-art performance on the ODinW-13 benchmark and ODinW-O, a newly introduced benchmark designed to assess class reappearance.
</p>
</div>
</div>
<!-- Centered Image -->
<div class="row justify-content-center mt-4 mb-5">
<div class="col-md-8 text-center">
<img src="[email protected]" class="img-fluid model-diagram" style="max-width: 80%; height: auto;">
</div>
</div>
</div>
</div>
<!-- Key Features - Light BG -->
<div class="content-section light-section">
<div class="container">
<div class="row justify-content-center mb-5">
<div class="col-md-10">
<h3 class="section-title text-center">Core Features & Innovations</h3>
<!-- First row -->
<div class="row mb-4 g-4 mt-4">
<!-- Problem Statement -->
<div class="col-md-6">
<div class="card feature-card">
<div class="card-body p-4">
<div class="text-center mb-3">
<i class="fas fa-exclamation-circle card-icon" style="color: #dc3545;"></i>
</div>
<h4 class="card-title text-center" style="color: #dc3545;">Problem Statement</h4>
<p class="card-text">
Vision-language detectors can incrementally learn new categories without losing zero-shot abilities, but existing methods rely on <strong>monolithic adaptation</strong>, merging all knowledge into one model.
This makes updating individual concepts difficult and can lead to knowledge interference and performance loss.
</p>
</div>
</div>
</div>
<!-- Modular Adaptation -->
<div class="col-md-6">
<div class="card feature-card">
<div class="card-body p-4">
<div class="text-center mb-3">
<i class="fas fa-code-branch card-icon" style="color: #198754;"></i>
</div>
<h4 class="card-title text-center" style="color: #198754;">Modular Adaptation</h4>
<p class="card-text">
<strong>DitHub</strong> introduces a modular approach by maintaining a library of specialized detection modules. Instead of integrating new knowledge into a single model, this approach enables efficient retrieval, fusion, and adaptation, ensuring flexible and scalable incremental learning.
</p>
</div>
</div>
</div>
</div>
<!-- Second row -->
<div class="row mb-4 g-4">
<!-- Specialized Modules -->
<div class="col-md-6">
<div class="card feature-card">
<div class="card-body p-4">
<div class="text-center mb-3">
<i class="fas fa-puzzle-piece card-icon" style="color: #0076B6;"></i>
</div>
<h4 class="card-title text-center" style="color: #0076B6">Specialized Modules</h4>
<p class="card-text">
Our experiments show that specialized modules improve performance significantly, with a state-of-the-art gain of <strong>+4.21 mAP</strong> in the Incremental Vision-Language Object Detection paradigm.
</p>
</div>
</div>
</div>
<!-- Training-free Unlearning -->
<div class="col-md-6">
<div class="card feature-card">
<div class="card-body p-4">
<div class="text-center mb-3">
<i class="fas fa-eraser card-icon" style="color: #ffc107;"></i>
</div>
<h4 class="card-title text-center" style="color: #ffc107;">Training-free Unlearning</h4>
<p class="card-text">
DitHub allows <strong>selective knowledge removal</strong> without retraining, enabling the efficient elimination of specific classes by removing associated modules.
</p>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
<!-- Model Overview -->
<div class="content-section">
<div class="container">
<div class="row justify-content-center mb-5">
<div class="col-md-10">
<h3 class="section-title text-center">Model Overview</h3>
<!-- Image with effects -->
<div class="text-center mb-5 mt-4">
<img src="[email protected]" alt="DitHub Model Architecture" class="img-fluid model-diagram" style="max-width: 95%; height: auto;">
</div>
<!-- Further details -->
<p class="lead text-center mb-5">
DitHub creates modular class-specific detectors, handling both rare and common objects while offering a memory-efficient structure for scalable long-term use. By using low-rank adaptation (LoRA), we fine-tune specialized modules for each task, ensuring minimal interference. This modular design supports on-the-fly adaptation to new classes, making DitHub a versatile solution for evolving detection needs.
</p>
<!-- Features Section -->
<div class="row justify-content-center g-4">
<div class="col-md-4">
<div class="card feature-card">
<div class="card-body p-4 text-center">
<div class="mb-3">
<i class="fas fa-layer-group card-icon"></i>
</div>
<h5 class="card-title">Modular Adaptation</h5>
<p class="card-text">
Expand the library of specialized modules dynamically as new tasks arise, adapting to both common and rare object classes.
</p>
</div>
</div>
</div>
<div class="col-md-4">
<div class="card feature-card">
<div class="card-body p-4 text-center">
<div class="mb-3">
<i class="fas fa-memory card-icon"></i>
</div>
<h5 class="card-title">Memory Efficiency</h5>
<p class="card-text">
Employ a subset of shared learnable parameters across different adaptation modules to enhance efficiency without compromising performance.
</p>
</div>
</div>
</div>
<div class="col-md-4">
<div class="card feature-card">
<div class="card-body p-4 text-center">
<div class="mb-3">
<i class="fas fa-cogs card-icon"></i>
</div>
<h5 class="card-title">Flexible Inference</h5>
<p class="card-text">
Enable selective activation of specialized modules, allowing direct deployment for fine-grained adaptation to individual classes.
</p>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
<!-- Experimental Results -->
<div class="content-section light-section">
<div class="container">
<div class="row justify-content-center">
<div class="col-md-10">
<h3 class="section-title">Experimental Results</h3>
<h4 class="mt-5 mb-3 fw-bold">Performance on ODinW-13</h4>
<p class="lead mb-4">
We evaluate DitHub on the ODinW-13 benchmark, which consists of 13 sub-datasets spanning from traditional Pascal VOC to
more challenging domains with significant distribution shifts. Our method outperforms the main competitor, ZiRa, by
a substantial <span class="highlight-text">4.21 mAP points</span> on Avg. Notably, on the zero-shot MS COCO evaluation (ZCOCO), our approach outperforms ZiRa by
0.75 mAP, setting a new state-of-the-art in both incremental and zero-shot retention capabilities.
</p>
<div class="table-responsive mb-5">
<table class="table table-striped text-center table-results">
<thead>
<tr>
<th>Shots</th> <th>Method</th> <th>ZCOCO</th> <th>Avg</th> <th>Ae</th> <th>Aq</th> <th>Co</th> <th>Eg</th>
<th>Mu</th> <th>Pa</th> <th>Pv</th> <th>Pi</th> <th>Po</th> <th>Ra</th> <th>Sh</th> <th>Th</th> <th>Ve</th>
</tr>
</thead>
<tbody>
<tr>
<td>0</td> <td>G-Dino</td> <td>47.41</td> <td>46.80</td> <td>19.11</td> <td>20.82</td> <td>64.75</td> <td>59.98</td>
<td>25.34</td> <td>56.27</td> <td>54.80</td> <td>65.94</td> <td>22.13</td> <td>62.02</td> <td>32.85</td> <td>70.38</td> <td>57.07</td>
</tr>
<tr>
<td rowspan="2" style="vertical-align: middle; text-align: center;">Full</td> <td>ZiRA</td> <td>46.26</td> <td>57.98</td> <td>31.76</td> <td>47.35</td> <td class="best-result">71.77</td>
<td>64.74</td> <td>46.53</td> <td>62.66</td> <td>66.39</td> <td>71.00</td> <td>48.48</td> <td>63.03</td> <td>41.44</td> <td>76.13</td> <td>62.44</td>
</tr>
<tr>
<td class="best-result">DitHub</td> <td class="best-result">47.01</td> <td class="best-result">62.19</td> <td class="best-result">34.62</td> <td class="best-result">50.65</td>
<td>70.46</td> <td class="best-result">68.56</td> <td class="best-result">49.28</td> <td class="best-result">65.57</td> <td class="best-result">69.58</td> <td class="best-result">71.10</td>
<td class="best-result">56.65</td> <td class="best-result">70.88</td> <td class="best-result">52.82</td> <td class="best-result">79.30</td> <td class="best-result">68.18</td>
</tr>
</tbody>
</table>
</div>
<h4 class="mt-5 mb-3 fw-bold">Performance on ODinW-O</h4>
<p class="lead mb-4">
We introduce ODinW-O (Overlapped), a variant of ODinW-35 specifically designed to evaluate performance on classes that
reoccur across different domains. In this benchmark, DitHub secures a substantial improvement of <span class="highlight-text">4.75 mAP</span> on Avg and a <span class="highlight-text">2.08 mAP</span>
gain on ZCOCO over ZiRa. We attribute this success to our class-oriented modular design, which enables selective updates
to recurring concepts without overwriting knowledge associated with other classes.
</p>
<div class="text-center mb-5">
<div style="width: 75%; margin: auto;">
<div class="table-responsive">
<table class="table table-striped text-center table-results">
<thead>
<tr>
<th>Shots</th><th>Method</th> <th>ZCOCO</th> <th>Avg</th> <th>Ae</th> <th>Hw</th> <th>Pv</th> <th>Sd</th> <th>Th</th> <th>Ve</th>
</tr>
</thead>
<tbody>
<tr>
<td>0</td><td>G-Dino</td> <td>47.41</td> <td>53.15</td> <td>45.12</td> <td>67.54</td> <td>58.11</td> <td>25.84</td> <td>70.40</td> <td>51.87</td>
</tr>
<tr>
<td rowspan="2" style="vertical-align: middle; text-align: center;">Full</td>
<td>ZiRa</td> <td>44.43</td> <td>57.63</td> <td>39.92</td> <td>68.00</td> <td>64.90</td> <td class="best-result">46.26</td> <td>77.26</td> <td>49.47</td>
</tr>
<tr>
<td class="best-result">DitHub</td> <td class="best-result">46.51</td> <td class="best-result">62.38</td> <td class="best-result">53.35</td> <td class="best-result">71.07</td> <td class="best-result">71.01</td> <td>41.75</td> <td class="best-result">80.21</td> <td class="best-result">56.90</td>
</tr>
</tbody>
</table>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
<!-- Citation Section -->
<div class="content-section" style="padding-top: 0;">
<div class="container">
<div class="row justify-content-center">
<div class="col-md-10">
<h3 class="section-title">Citation</h3>
<p class="lead mb-3">
If you find DitHub useful for your research, please consider citing our paper:
</p>
<div class="card p-4 mb-4" style="border-radius: 15px; background-color: #f8f9fa; border: none; box-shadow: 0 5px 15px rgba(0, 0, 0, 0.05);">
<div style="overflow-x: auto;">
<pre style="background-color: transparent; border: none; padding: 10px; margin: 0; font-size: 0.8rem; line-height: 1.5; font-family: 'Courier New', Courier, monospace; color: #333;">
@article{cappellino2025dithub,
title={DitHub: A Modular Framework for Incremental Open-Vocabulary Object Detection},
author={Cappellino, Chiara and Mancusi, Gianluca and Mosconi, Matteo and Porrello, Angelo and Calderara, Simone and Cucchiara, Rita},
journal={arXiv preprint arXiv:2503.09271},
year={2025}
}
</pre>
</div>
</div>
</div>
</div>
</div>
</div>
<!-- Footer -->
<footer class="py-4 bg-dark text-white text-center">
<div class="container">
<p class="mb-0">DitHub: Incremental Open-Vocabulary Object Detection</p>
</div>
</footer>
<script src="https://cdnjs.cloudflare.com/ajax/libs/bootstrap/5.3.2/js/bootstrap.bundle.min.js"></script>
</body>
</html>