|
| 1 | +<!DOCTYPE html> |
| 2 | +<html> |
| 3 | +<head> |
| 4 | + <meta charset="utf-8"> |
| 5 | + <meta name="description" |
| 6 | + content="Versatile and Guided Flow Matching Model of Human Pose."> |
| 7 | + <meta name="keywords" content="PoseD-Flow"> |
| 8 | + <meta name="viewport" content="width=device-width, initial-scale=1"> |
| 9 | + <title>PoseD-Flow: Versatile and Guided Flow Matching Model of Human Pose</title> |
| 10 | + |
| 11 | + <link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" |
| 12 | + rel="stylesheet"> |
| 13 | + |
| 14 | + <link rel="stylesheet" href="./static/css/bulma.min.css"> |
| 15 | + <link rel="stylesheet" href="./static/css/bulma-carousel.min.css"> |
| 16 | + <link rel="stylesheet" href="./static/css/bulma-slider.min.css"> |
| 17 | + <link rel="stylesheet" href="./static/css/fontawesome.all.min.css"> |
| 18 | + <link rel="stylesheet" |
| 19 | + href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css"> |
| 20 | + <link rel="stylesheet" href="./static/css/index.css"> |
| 21 | + |
| 22 | + <script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script> |
| 23 | + <script defer src="./static/js/fontawesome.all.min.js"></script> |
| 24 | + <script src="./static/js/bulma-carousel.min.js"></script> |
| 25 | + <script src="./static/js/bulma-slider.min.js"></script> |
| 26 | + <script src="./static/js/index.js"></script> |
| 27 | +</head> |
| 28 | +<body> |
| 29 | + |
| 30 | +<section class="hero"> |
| 31 | + <div class="hero-body" style="padding-bottom: 0.5rem;"> |
| 32 | + <div class="container is-max-desktop"> |
| 33 | + <div class="columns is-centered"> |
| 34 | + <div class="column has-text-centered"> |
| 35 | + <h1 class="title is-1 publication-title" style="margin-bottom: 0.45rem;">PoseD-Flow: Versatile and Guided Flow Matching Model of Human Pose</h1> |
| 36 | + <div class="is-size-4 publication-authors"> |
| 37 | + <span class="author-block">CVPR 2026</span> |
| 38 | + </div> |
| 39 | + <div class="is-size-4 publication-authors"> |
| 40 | + <span class="author-block"> |
| 41 | + <a href="https://scholar.google.com/citations?user=9DiuOycAAAAJ">Jebastin Nadar</a>,</span> |
| 42 | + <span class="author-block"> |
| 43 | + <a href="https://www.simofoti.com/">Simone Foti</a>,</span> |
| 44 | + <span class="author-block"> |
| 45 | + <a href="https://tolgabirdal.github.io/">Tolga Birdal</a> |
| 46 | + </span> |
| 47 | + </div> |
| 48 | + <div class="is-size-4 publication-authors"> |
| 49 | + <span class="author-block">Imperial College London</span> |
| 50 | + </div> |
| 51 | + |
| 52 | + <div class="column has-text-centered"> |
| 53 | + <div class="publication-links"> |
| 54 | + <span class="link-block"> |
| 55 | + <a href="" |
| 56 | + class="external-link button is-normal is-rounded is-dark"> |
| 57 | + <span class="icon"> |
| 58 | + <i class="fas fa-file-pdf"></i> |
| 59 | + </span> |
| 60 | + <span>Paper</span> |
| 61 | + </a> |
| 62 | + </span> |
| 63 | + <span class="link-block"> |
| 64 | + <a href="" |
| 65 | + class="external-link button is-normal is-rounded is-dark"> |
| 66 | + <span class="icon"> |
| 67 | + <i class="ai ai-arxiv"></i> |
| 68 | + </span> |
| 69 | + <span>arXiv</span> |
| 70 | + </a> |
| 71 | + </span> |
| 72 | + <span class="link-block"> |
| 73 | + <a href="https://github.com/circle-group/PoseDFlow" |
| 74 | + class="external-link button is-normal is-rounded is-dark"> |
| 75 | + <span class="icon"> |
| 76 | + <i class="fab fa-github"></i> |
| 77 | + </span> |
| 78 | + <span>Code</span> |
| 79 | + </a> |
| 80 | + </span> |
| 81 | + </div> |
| 82 | + </div> |
| 83 | + </div> |
| 84 | + </div> |
| 85 | + </div> |
| 86 | + </div> |
| 87 | +</section> |
| 88 | + |
| 89 | +<section class="section" style="padding-top: 0.5rem; padding-bottom: 1.75rem;"> |
| 90 | + <div class="container" style="max-width: var(--overview-max-width, 1000px); width: 100%;"> |
| 91 | + <div class="columns is-centered has-text-centered"> |
| 92 | + <div class="column is-full"> |
| 93 | + <div class="content has-text-justified" style="border: 1px solid #d9d9d9; border-radius: 12px; padding: 2rem; background: #fff;"> |
| 94 | + <h2 class="title is-3 has-text-centered">Overview</h2> |
| 95 | + <p style="font-size: 1.1rem;"> |
| 96 | + Generative pose priors have recently emerged as a powerful tool for inference under occlusion or noise. Yet today's strongest generative paradigm, flow matching, remains unused for human pose due to two fundamental barriers: the absence of a pre-trained flow prior and the non-Euclidean nature of articulated poses. We overcome both by introducing <strong>PoseD-Flow</strong>, a novel framework to unify Riemannian Flow Matching (RFM) with training-free guidance for 3D human pose recovery. PoseD-Flow is composed of two contributions: (i) <strong>PoseRFM</strong>, the first RFM model of human pose, defined directly on the product manifold of joint rotations, and (ii) <strong>Riemannian D-Flow</strong>, a principled guidance mechanism that, by differentiating through its ODE sampling dynamics, conditions PoseRFM at inference without any task-specific training. Our theoretical analysis shows that the induced dynamics are shaped by data covariance and manifold curvature, yielding a bias toward realistic poses. Across pose completion, denoising, and inverse kinematics, PoseD-Flow establishes new state of the art, particularly under noise, occlusion, and partial observations. |
| 97 | + </p> |
| 98 | + <img id="teaser" src="./static/images/teaser_v2.png" alt="Teaser image" style="display: block; width: 100%; height: auto; margin: 1rem auto 0.75rem;"> |
| 99 | + <p style="font-size: 1.05rem; text-align: center;"><strong>PoseD-Flow</strong> framework: (top) <strong>PoseRFM</strong>, a robust human pose prior defined on the product manifold of joints using <em>Riemannian Flow Matching;</em> (bottom): <strong>Riemannian D-Flow</strong>, a flexible, geometry-aware inversion technique for flow models. Together, they provide a novel approach to solving inverse problems in human pose, achieving results competitive with SotA diffusion models.</p> |
| 100 | + </div> |
| 101 | + </div> |
| 102 | + </div> |
| 103 | + </div> |
| 104 | +</section> |
| 105 | + |
| 106 | +<section class="section" style="padding-top: 1.75rem;"> |
| 107 | + <div class="container is-max-desktop has-text-centered" style="margin-bottom: 1rem;"> |
| 108 | + <h2 class="title is-3">Results</h2> |
| 109 | + </div> |
| 110 | + <div style="width: 100%; height: 2px; background: linear-gradient(to right, rgba(0, 0, 0, 0), rgba(0, 0, 0, 1), rgba(0, 0, 0, 0));"></div> |
| 111 | + <div class="container has-text-centered" style="max-width: var(--results-max-width, 1000px); width: 100%; margin-top: 3rem;"> |
| 112 | + <div style="margin-bottom: 3rem; border: 1px solid #d9d9d9; border-radius: 12px; padding: 1.5rem; background: #fff;"> |
| 113 | + <h3 class="title is-3">Pose Generation</h3> |
| 114 | + <p style="font-size: 1.1rem;">We visualize the sampling trajectory for unconditional pose generation, where an initial random noise sample is gradually refined as the ODE solver advances from N=0 to N=100 function evaluations.</p> |
| 115 | + <video id="pose_gen" autoplay controls muted loop playsinline height="100%" style="display: block; width: 100%; max-width: 960px; margin: 0 auto;"> |
| 116 | + <source src="./static/videos/pose_gen.webm" |
| 117 | + type="video/webm"> |
| 118 | + </video> |
| 119 | + </div> |
| 120 | + <div style="margin-bottom: 3rem; border: 1px solid #d9d9d9; border-radius: 12px; padding: 1.5rem; background: #fff;"> |
| 121 | + <h3 class="title is-3">Pose Completion</h3> |
| 122 | + <p style="font-size: 1.1rem;">Pose completion with <span style="color: #D1AECCFF; font-weight: 600;">occluded joints</span>. The animation shows the evolution of the final pose from N=0 to N=200 iterations optimized by PoseD-Flow.</p> |
| 123 | + <video id="pose_comp" autoplay controls muted loop playsinline height="100%" style="display: block; width: 100%; max-width: 960px; margin: 0 auto;"> |
| 124 | + <source src="./static/videos/pose_comp.webm" |
| 125 | + type="video/webm"> |
| 126 | + </video> |
| 127 | + <div style="width: 100%; max-width: 960px; margin: 0.8rem auto 0; display: flex; justify-content: center; align-items: center; gap: 8.4rem; font-size: 1.25rem; font-weight: 700; line-height: 1.2;"> |
| 128 | + <span style="text-align: center;">Ground Truth</span> |
| 129 | + <span style="text-align: center;">Completed Pose</span> |
| 130 | + </div> |
| 131 | + </div> |
| 132 | + <div style="margin-bottom: 3rem; border: 1px solid #d9d9d9; border-radius: 12px; padding: 1.5rem; background: #fff;"> |
| 133 | + <h3 class="title is-3">Motion Denoising</h3> |
| 134 | + <p style="font-size: 1.1rem;">Even though our method is not explicitly trained for motion tasks, it is still able to recover plausible motion from noisy 3D observations.</p> |
| 135 | + <video id="motion_denoise" autoplay controls muted loop playsinline height="100%" style="display: block; width: 100%; max-width: 960px; margin: 0 auto;"> |
| 136 | + <source src="./static/videos/motion_denoise.webm" |
| 137 | + type="video/webm"> |
| 138 | + </video> |
| 139 | + <div style="width: 100%; max-width: 960px; margin: 0.8rem auto 0; display: flex; justify-content: space-between; align-items: center; font-size: 1.25rem; font-weight: 700; line-height: 1.2;"> |
| 140 | + <span style="text-align: center; width: 33.33%;">Ground Truth</span> |
| 141 | + <span style="text-align: center; width: 33.33%;">Noisy Input</span> |
| 142 | + <span style="text-align: center; width: 33.33%;">Denoised Output</span> |
| 143 | + </div> |
| 144 | + </div> |
| 145 | + <div style="border: 1px solid #d9d9d9; border-radius: 12px; padding: 1.5rem; background: #fff;"> |
| 146 | + <h3 class="title is-3">Human Mesh Recovery</h3> |
| 147 | + <p style="font-size: 1.1rem;">HMR fitting results on some in-the-wild images from 3DPW.</p> |
| 148 | + <div style="width: 100%; max-width: 960px; margin: 0.8rem auto; display: flex; justify-content: space-between; gap: 1.5rem;"> |
| 149 | + <div style="flex: 1; display: grid; grid-template-columns: repeat(3, 1fr); gap: 0.0rem;"> |
| 150 | + <img src="./static/images/ours_00180.png" alt="From Scratch 1" style="width: 100%; height: 95%; display: block;"> |
| 151 | + <img src="./static/images/ours_00422_cropped.png" alt="From Scratch 2" style="width: 100%; height: 95%; display: block;"> |
| 152 | + <img src="./static/images/ours_00427.png" alt="From Scratch 3" style="width: 100%; height: 95%; display: block;"> |
| 153 | + </div> |
| 154 | + <div style="flex: 1; display: grid; grid-template-columns: repeat(3, 1fr); gap: 0.0rem;"> |
| 155 | + <img src="./static/images/ours_00682.png" alt="Init from CLIFF 1" style="width: 100%; height: 95%; display: block;"> |
| 156 | + <img src="./static/images/ours_00305_cropped.png" alt="Init from CLIFF 2" style="width: 100%; height: 95%; display: block;"> |
| 157 | + <img src="./static/images/ours_00724_cropped.png" alt="Init from CLIFF 3" style="width: 100%; height: 95%; display: block;"> |
| 158 | + </div> |
| 159 | + </div> |
| 160 | + <div style="width: 100%; max-width: 960px; margin: 0.8rem auto 0; display: flex; justify-content: space-between; align-items: center; font-size: 1.25rem; font-weight: 700; line-height: 1.2;"> |
| 161 | + <span style="text-align: center; width: 50%;">From Scratch</span> |
| 162 | + <span style="text-align: center; width: 50%;">Init. from CLIFF</span> |
| 163 | + </div> |
| 164 | + </div> |
| 165 | + </div> |
| 166 | +</section> |
| 167 | + |
| 168 | +<section class="section" id="BibTeX" style="padding-top: 0.5rem; padding-bottom: 1.75rem;"> |
| 169 | + <div class="container is-max-desktop content"> |
| 170 | + <h2 class="title">BibTeX</h2> |
| 171 | + <div style="position: relative;"> |
| 172 | + <button id="copy-bibtex" class="button is-small is-light" style="position: absolute; top: 0.5rem; right: 0.5rem; z-index: 1;" aria-label="Copy BibTeX" title="Copy BibTeX"> |
| 173 | + <span class="icon is-small"> |
| 174 | + <i class="fas fa-copy"></i> |
| 175 | + </span> |
| 176 | + </button> |
| 177 | + <span id="copy-bibtex-popup" style="display: none; position: absolute; top: 0.5rem; right: 3rem; background: rgba(54, 54, 54, 0.95); color: #fff; font-size: 0.75rem; padding: 0.2rem 0.45rem; border-radius: 4px; z-index: 2;">Copied</span> |
| 178 | + <pre></pre> |
| 179 | + </div> |
| 180 | + </div> |
| 181 | +</section> |
| 182 | + |
| 183 | + |
| 184 | +<footer class="footer"> |
| 185 | + <div class="container"> |
| 186 | + <div class="columns is-centered"> |
| 187 | + <div class="column is-8"> |
| 188 | + <div class="content"> |
| 189 | + <p> |
| 190 | + This website is licensed under a <a rel="license" |
| 191 | + href="http://creativecommons.org/licenses/by-sa/4.0/">Creative |
| 192 | + Commons Attribution-ShareAlike 4.0 International License</a>. The template is borrowed from <a href="https://github.com/nerfies/nerfies.github.io">Nerfies</a>. |
| 193 | + </p> |
| 194 | + </div> |
| 195 | + </div> |
| 196 | + </div> |
| 197 | + </div> |
| 198 | +</footer> |
| 199 | + |
| 200 | +<script> |
| 201 | + document.getElementById('copy-bibtex').addEventListener('click', async function () { |
| 202 | + const bibtex = document.getElementById('bibtex-code').innerText; |
| 203 | + await navigator.clipboard.writeText(bibtex); |
| 204 | + const popup = document.getElementById('copy-bibtex-popup'); |
| 205 | + popup.style.display = 'inline-block'; |
| 206 | + setTimeout(function () { |
| 207 | + popup.style.display = 'none'; |
| 208 | + }, 2000); |
| 209 | + }); |
| 210 | +</script> |
| 211 | + |
| 212 | +</body> |
| 213 | +</html> |
0 commit comments