TimePLE-MAC.github.io/index.html at main · MAC-AutoML/TimePLE-MAC.github.io · GitHub

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
<!DOCTYPE html>
<html lang="en">
<head>
  <meta charset="utf-8">
  <meta name="viewport" content="width=device-width, initial-scale=1">
  <meta name="title" content="TimePLE: Rethinking Temporal Representation for Video Temporal Grounding">
  <meta name="description" content="TimePLE reformulates video temporal grounding as interval-native prediction over a canonical position-duration space.">
  <meta name="keywords" content="TimePLE, video temporal grounding, temporal representation, multimodal language model, interval prediction">
  <meta name="author" content="Yuhui Zeng, Xinyu Mao, Xiaokun Liu, Xin Tao, Pengfei Wan, Jinfa Huang, Jiayi Ji, Xiawu Zheng">
  <meta name="robots" content="index, follow">
  <meta property="og:type" content="article">
  <meta property="og:title" content="TimePLE: Rethinking Temporal Representation for Video Temporal Grounding">
  <meta property="og:description" content="From endpoint prediction to interval-native temporal grounding.">
  <meta property="og:image" content="static/images/teaser_final.png">
  <meta name="twitter:card" content="summary_large_image">
  <meta name="citation_title" content="TimePLE: Rethinking Temporal Representation for Video Temporal Grounding">
  <meta name="citation_author" content="Zeng, Yuhui">
  <meta name="citation_author" content="Mao, Xinyu">
  <meta name="citation_author" content="Liu, Xiaokun">
  <meta name="citation_author" content="Tao, Xin">
  <meta name="citation_author" content="Wan, Pengfei">
  <meta name="citation_author" content="Huang, Jinfa">
  <meta name="citation_author" content="Ji, Jiayi">
  <meta name="citation_author" content="Zheng, Xiawu">
  <meta name="citation_publication_date" content="2026">
  <meta name="citation_pdf_url" content="static/pdfs/timeple.pdf">
  <meta name="theme-color" content="#146c52">
  <title>TimePLE | Interval-Native Video Temporal Grounding</title>
  <link rel="icon" type="image/png" href="static/images/favicon.png">
  <link rel="stylesheet" href="static/css/bulma.min.css">
  <link rel="stylesheet" href="static/css/index.css">
  <link rel="preload" href="static/css/bulma-carousel.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
  <link rel="preload" href="static/css/bulma-slider.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
  <link rel="preload" href="static/css/fontawesome.all.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
  <noscript><link rel="stylesheet" href="static/css/bulma-carousel.min.css"><link rel="stylesheet" href="static/css/fontawesome.all.min.css"></noscript>
  <script defer src="static/js/fontawesome.all.min.js"></script>
  <script defer src="static/js/bulma-carousel.min.js"></script>
  <script defer src="static/js/bulma-slider.min.js"></script>
  <script defer src="static/js/index.js"></script>
</head>
<body>
  <button class="scroll-to-top" onclick="scrollToTop()" title="Scroll to top" aria-label="Scroll to top"><i class="fas fa-chevron-up"></i></button>

  <!-- Template component: publication hero -->
  <section class="hero publication-header" id="hero">
    <div class="hero-body">
      <div class="container is-max-desktop">
        <div class="columns is-centered">
          <div class="column has-text-centered">
            <img class="publication-logo" src="static/images/timeple-logo.png" alt="TimePLE">
            <h1 class="title is-1 publication-title">Rethinking Temporal Representation<br>for Video Temporal Grounding</h1>
            <p class="publication-tagline">From endpoint prediction to interval-native grounding.</p>
            <div class="is-size-5 publication-authors">
              <span class="author-block">Yuhui Zeng<sup>1,4,*†</sup>,</span>
              <span class="author-block">Xinyu Mao<sup>2,4,*†</sup>,</span>
              <span class="author-block">Xiaokun Liu<sup>4,‡✉</sup>,</span>
              <span class="author-block">Xin Tao<sup>4</sup>,</span>
              <span class="author-block">Pengfei Wan<sup>4</sup>,</span>
              <span class="author-block">Jinfa Huang<sup>3</sup>,</span>
              <span class="author-block">Jiayi Ji<sup>1</sup>,</span>
              <span class="author-block">Xiawu Zheng<sup>1,✉</sup></span>
            </div>
            <div class="is-size-6 publication-authors affiliations">
              <span class="author-block"><sup>1</sup>Xiamen University</span>
              <span class="author-block"><sup>2</sup>The Chinese University of Hong Kong</span>
              <span class="author-block"><sup>3</sup>University of Rochester</span>
              <span class="author-block"><sup>4</sup>Kling Team, Kuaishou Technology</span>
              <span class="eql-cntrb"><small><br><sup>*</sup>Equal contribution &nbsp; <sup>†</sup>Work conducted during internship &nbsp; <sup>‡</sup>Project leader</small></span>
            </div>
            <div class="column has-text-centered">
              <div class="publication-links">
                <span class="link-block"><a href="static/pdfs/timeple.pdf" class="external-link button is-normal is-rounded is-dark"><span class="icon"><i class="fas fa-file-pdf"></i></span><span>Paper</span></a></span>
                <span class="link-block"><a href="#BibTeX" class="external-link button is-normal is-rounded is-dark"><span class="icon"><i class="fas fa-quote-right"></i></span><span>BibTeX</span></a></span>
                <span class="link-block"><span class="button is-normal is-rounded is-light coming-soon"><span class="icon"><i class="fab fa-github"></i></span><span>Code Soon</span></span></span>
                <span class="link-block"><span class="button is-normal is-rounded is-light coming-soon"><span class="icon"><i class="fas fa-cube"></i></span><span>Models Soon</span></span></span>
                <span class="link-block"><span class="button is-normal is-rounded is-light coming-soon"><span class="icon"><i class="fas fa-database"></i></span><span>Data Soon</span></span></span>
              </div>
            </div>
          </div>
        </div>
      </div>
    </div>
  </section>

  <!-- Template component: teaser -->
  <section class="hero teaser">
    <div class="container is-max-desktop">
      <div class="hero-body">
        <figure class="publication-banner"><img src="static/images/teaser_final.png" alt="Endpoint prediction versus TimePLE interval-native grounding"></figure>
        <h2 class="subtitle has-text-centered teaser-caption"><strong>A temporal moment is not two independent timestamps.</strong> TimePLE predicts one joint distribution over valid temporal intervals, making placement, duration, and interval-level similarity explicit.</h2>
      </div>
    </div>
  </section>

  <!-- Template section: Overview and motivation -->
  <section class="section hero is-light" id="overview">
    <div class="container is-max-desktop">
      <div class="overview-heading"><p class="section-kicker">Overview · From Endpoints to Intervals</p><h2 class="title is-3">Model the moment as an interval.</h2></div>
      <div class="columns is-variable is-6">
        <div class="column"><article class="content-card endpoint-card"><p class="subsection-label">Endpoint prediction</p><h3>Two decisions fragment one moment.</h3><div class="timeline"><span class="event"></span><i class="start">start</i><i class="end">end</i></div><p>The event interval is assembled afterward; duration and interval-level similarity remain implicit.</p></article></div>
        <div class="column"><article class="content-card interval-card"><p class="subsection-label">Interval-native prediction</p><h3>One coordinate represents the complete span.</h3><div class="mapping"><b>I = [s,e]</b><span>→</span><b>(u,v)</b><small>placement · duration</small></div><p>Every canonical coordinate maps to a valid interval and directly supports interval-level prediction.</p></article></div>
      </div>
    </div>
  </section>

  <section class="section hero is-light" id="method">
    <div class="container is-max-desktop">
      <div class="columns is-centered"><div class="column"><p class="section-kicker">§2 · TimePLE</p><h2 class="title is-3">A lightweight interval codec for VLMs.</h2><p class="section-lead">TimePLE preserves the autoregressive language interface while giving temporal grounding a native interval geometry.</p></div></div>
      <figure class="publication-body framework-figure"><img src="static/images/main_timeple.png" alt="TimePLE architecture"><figcaption>Input temporal anchors and output spans share the same canonical interval space.</figcaption></figure>
      <div class="columns method-columns">
        <div class="column"><article class="method-card"><span>§2.1</span><h3>Canonical Position–Duration Space</h3><p><b>Why.</b> Start–end coordinates occupy a constrained triangle and do not expose duration directly.</p><div class="equation">v = e − s<br>u = s / (1 − v)</div><p><b>Advantage.</b> Its inverse maps every square coordinate to a valid interval.</p></article></div>
        <div class="column"><article class="method-card"><span>§2.2</span><h3>One Span Token, One Distribution</h3><p><b>Why.</b> A complete moment should be supported by one coherent latent prediction.</p><div class="token-flow"><b>&lt;|TIMESPAN|&gt;</b><i>→</i><b>p(u,v)</b><i>→</i><b>𝔼[u,v]</b></div><p><b>Advantage.</b> Joint decoding retains uncertainty before producing continuous boundaries.</p></article></div>
        <div class="column"><article class="method-card"><span>§2.3</span><h3>Duration-Conditioned Refinement</h3><p><b>Why.</b> Short events suffer larger IoU drops from the same coordinate error.</p><div class="refinement"><span></span><small>shorter span → stronger correction</small></div><p><b>Advantage.</b> A bounded residual improves precision beyond grid expectation.</p></article></div>
      </div>
    </div>
  </section>

  <section class="section results-section" id="experiments">
    <div class="container is-max-desktop">
      <p class="section-kicker">§3 · Experimental Evidence</p>
      <h2 class="title is-3">Interval-native grounding improves temporal localization.</h2>
      <p class="section-lead">Across four VTG benchmarks, TimePLE achieves the best average mIoU among the compared temporal-output paradigms, with its clearest advantages on short and medium events.</p>

      <div class="experiment-subsection">
        <div class="experiment-heading"><div><p class="subsection-label">§3.1 Main Results</p><h3>Comparison across four VTG benchmarks.</h3></div><div class="main-score"><strong>58.9</strong><span>Avg. mIoU</span></div></div>
        <a class="main-table-figure" href="static/images/main_results_table.png" target="_blank" title="Open the full-resolution table">
          <img src="static/images/main_results_table.png" alt="Main results on four video temporal grounding benchmarks">
        </a>
        <p class="figure-note">TimePLE-8B achieves 57.2 on Charades-STA, 49.6 on ActivityNet-Captions, 65.3 on QVHighlights, and 63.4 on Charades-TimePLE. Click the table to inspect the full-resolution results.</p>
      </div>

      <div class="experiment-subsection duration-results">
        <div><p class="subsection-label">§3.2 Duration-Stratified Performance</p><h3>Fine-grained localization matters most for shorter moments.</h3><p>Small temporal errors cause a larger loss of interval overlap for short and medium events. TimePLE improves these duration regimes while remaining competitive on long moments.</p></div>
        <div class="duration-highlights"><div><span>Charades-STA</span><strong>55.5</strong><small>Short</small><strong>62.9</strong><small>Medium</small></div><div><span>Charades-TimePLE</span><strong>60.8</strong><small>Short</small><strong>71.1</strong><small>Medium</small></div></div>
      </div>
    </div>
  </section>

  <section class="hero is-small analysis-section" id="analysis">
    <div class="hero-body">
      <div class="container is-max-desktop">
        <p class="section-kicker">§4 · Representation-Level Analysis</p>
        <h2 class="title is-3">From endpoint-wise decisions to interval-level representations.</h2>
        <p class="section-lead">We examine whether changing the prediction object from two endpoints to one complete interval also changes the model's underlying temporal representation.</p>

        <figure class="publication-body framework-figure analysis-main-figure">
          <img src="static/images/representation_analysis_single_col.png" alt="Single-column representation analysis comparing temporal attribution and interval likelihood landscapes">
          <figcaption><b>Representation-level analysis.</b> (a) Gradient × Input attribution compares Timestamp-Text start–end targets with TimePLE position–duration targets. Dashed lines mark the ground-truth boundaries. (b) Timestamp-Text likelihood is reconstructed over valid start–end candidates, while TimePLE directly predicts a joint distribution in the canonical interval square. Crosses and diamonds denote ground truth and prediction.</figcaption>
        </figure>

        <div class="additional-analysis">
          <div class="additional-heading"><p class="subsection-label">Additional Qualitative Cases</p><h3>More representation-level comparisons.</h3></div>
          <figure class="additional-figure"><div class="figure-label">Additional temporal evidence attribution cases</div><img src="static/images/additional_attribution_cases.png" alt="Additional temporal evidence attribution cases" loading="lazy"><figcaption>Timestamp-Text uses the ground-truth start/end token likelihoods; TimePLE uses the ground-truth position/duration marginal likelihoods. Dashed lines mark the target interval.</figcaption></figure>
          <figure class="additional-figure"><div class="figure-label">Additional interval likelihood landscapes</div><img src="static/images/additional_landscape_cases.png" alt="Additional interval likelihood landscapes" loading="lazy"><figcaption>Timestamp-Text landscapes are reconstructed from 528 valid boundary candidates. TimePLE directly outputs the joint distribution over its canonical interval square.</figcaption></figure>
        </div>
      </div>
    </div>
  </section>

  <section class="section hero is-light data-section" id="data">
    <div class="container is-max-desktop">
      <div class="columns is-vcentered is-variable is-5"><div class="column is-one-third data-copy"><p class="section-kicker">Data & Benchmark Quality</p><h2 class="title is-3">Data Curation Pipeline</h2><p>A multi-model curation and human-verification pipeline constructs grounded training data and corrects noisy benchmark annotations.</p><div class="data-stats"><div><strong>90K</strong><span>scale grounded samples</span></div><div><strong>3K</strong><span>scale corrected annotations</span></div></div></div><div class="column is-two-thirds"><figure class="publication-body data-figure"><img src="static/images/Data_Curation_Pipeline.png" alt="Data curation and benchmark correction pipeline"></figure></div></div>
    </div>
  </section>

  <section class="section" id="full-results">
    <div class="container is-max-desktop">
      <p class="section-kicker">Full Results · Backbone Scaling</p>
      <h2 class="title is-3">Interval-native grounding scales across VLM backbones.</h2>
      <p class="section-lead">For every backbone, TimePLE and Timestamp-SFT use the same training data and video input configuration. The temporal prediction interface is the only change.</p>
      <div class="table-container scaling-table-wrap">
        <table class="table is-fullwidth result-table scaling-table">
          <thead><tr><th>Backbone</th><th>Method</th><th>C-STA</th><th>A-Net</th><th>QVH</th><th>C-TimePLE</th><th>Avg.</th></tr></thead>
          <tbody>
            <tr><td rowspan="2">LLaVA-OV-0.5B</td><td>Timestamp-SFT</td><td>23.8</td><td>10.8</td><td>6.5</td><td>24.7</td><td>16.5</td></tr>
            <tr class="timeple-row"><td>TimePLE</td><td><b>40.2</b> <small>+16.4</small></td><td><b>33.2</b> <small>+22.4</small></td><td><b>32.3</b> <small>+25.8</small></td><td><b>42.3</b> <small>+17.6</small></td><td><b>37.0</b> <small>+20.6</small></td></tr>
            <tr class="group-start"><td rowspan="2">Qwen2.5-VL-3B</td><td>Timestamp-SFT</td><td>51.2</td><td>41.6</td><td>59.3</td><td>57.8</td><td>52.5</td></tr>
            <tr class="timeple-row"><td>TimePLE</td><td><b>52.3</b> <small>+1.1</small></td><td><b>43.3</b> <small>+1.7</small></td><td><b>59.4</b> <small>+0.1</small></td><td><b>58.5</b> <small>+0.7</small></td><td><b>53.4</b> <small>+0.9</small></td></tr>
            <tr class="group-start"><td rowspan="2">Qwen3-VL-8B</td><td>Timestamp-SFT</td><td>51.6</td><td>47.3</td><td>63.6</td><td>56.7</td><td>54.8</td></tr>
            <tr class="timeple-row"><td>TimePLE</td><td><b>57.2</b> <small>+5.6</small></td><td><b>49.6</b> <small>+2.3</small></td><td><b>65.3</b> <small>+1.7</small></td><td><b>63.4</b> <small>+6.7</small></td><td><b>58.9</b> <small>+4.1</small></td></tr>
          </tbody>
        </table>
      </div>
      <p class="table-note">All values are mIoU. Green rows use TimePLE's interval-native representation; <span class="gain-note">+x.x</span> reports the absolute gain over the matched Timestamp-SFT baseline. C-STA, A-Net, QVH, and C-TimePLE denote Charades-STA, ActivityNet-Captions, QVHighlights, and Charades-TimePLE.</p>
    </div>
  </section>

  <section class="section hero is-light" id="resources">
    <div class="container is-max-desktop"><div class="columns is-centered has-text-centered"><div class="column"><p class="section-kicker">Resources</p><h2 class="title is-3">Read, reproduce, and build on TimePLE.</h2></div></div><div class="columns resource-columns"><div class="column"><a class="resource-card" href="static/pdfs/timeple.pdf"><i class="fas fa-file-pdf"></i><h3>Paper</h3><p>Full paper and appendix</p></a></div><div class="column"><div class="resource-card disabled"><i class="fab fa-github"></i><h3>Code</h3><p>Coming soon</p></div></div><div class="column"><div class="resource-card disabled"><i class="fas fa-cube"></i><h3>Models</h3><p>Coming soon</p></div></div><div class="column"><div class="resource-card disabled"><i class="fas fa-database"></i><h3>Data</h3><p>Coming soon</p></div></div></div></div>
  </section>

  <!-- Template component: BibTeX -->
  <section class="section" id="BibTeX">
    <div class="container is-max-desktop content"><div class="bibtex-header"><h2 class="title">BibTeX</h2><button class="copy-bibtex-btn" onclick="copyBibTeX()" title="Copy BibTeX"><i class="fas fa-copy"></i><span class="copy-text">Copy</span></button></div><pre id="bibtex-code"><code>@article{zeng2026timeple,
  title   = {TimePLE: Rethinking Temporal Representation for Video Temporal Grounding},
  author  = {Zeng, Yuhui and Mao, Xinyu and Liu, Xiaokun and Tao, Xin and Wan, Pengfei and Huang, Jinfa and Ji, Jiayi and Zheng, Xiawu},
  journal = {arXiv preprint},
  year    = {2026}
}</code></pre><p class="table-note">Provisional citation—update the venue and arXiv identifier when public.</p></div>
  </section>

  <!-- Template component: footer -->
  <footer class="footer"><div class="container"><div class="columns is-centered"><div class="column is-8"><div class="content has-text-centered"><img class="footer-logo" src="static/images/timeple-logo.png" alt="TimePLE"><p>From endpoint prediction to interval-native temporal grounding.</p><p class="footer-credit">This page was built using the <a href="https://github.com/eliahuhorwitz/Academic-project-page-template">Academic Project Page Template</a>, adopted from the <a href="https://nerfies.github.io">Nerfies</a> project page.</p></div></div></div></div></footer>
</body>
</html>