<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title>Transformer - Tag - 堂堂一跑堂</title><link>https://spacetop.win/tags/transformer/</link><description>Transformer - Tag - 堂堂一跑堂</description><generator>Hugo -- gohugo.io</generator><language>zh-CN</language><managingEditor>kingcopper@whu.edu.cn (WangTong)</managingEditor><webMaster>kingcopper@whu.edu.cn (WangTong)</webMaster><lastBuildDate>Mon, 01 Jun 2026 12:00:00 +0800</lastBuildDate><atom:link href="https://spacetop.win/tags/transformer/" rel="self" type="application/rss+xml"/><item><title>SkySense V2：统一多模态遥感基础模型，一个Backbone搞定RGB/SAR/多光谱</title><link>https://spacetop.win/2026/06/20260601_220000_skysense_v2_multimodal/</link><pubDate>Mon, 01 Jun 2026 12:00:00 +0800</pubDate><author><name>WangTong</name></author><guid>https://spacetop.win/2026/06/20260601_220000_skysense_v2_multimodal/</guid><description><![CDATA[<h1 id="skysense-v2统一多模态遥感基础模型一个backbone搞定rgbsar多光谱" class="headerLink">
    <a href="#skysense-v2%e7%bb%9f%e4%b8%80%e5%a4%9a%e6%a8%a1%e6%80%81%e9%81%a5%e6%84%9f%e5%9f%ba%e7%a1%80%e6%a8%a1%e5%9e%8b%e4%b8%80%e4%b8%aabackbone%e6%90%9e%e5%ae%9argbsar%e5%a4%9a%e5%85%89%e8%b0%b1" class="header-mark"></a>SkySense V2：统一多模态遥感基础模型，一个Backbone搞定RGB/SAR/多光谱</h1><blockquote>
  <p><strong>论文解读</strong> | ICCV 2025 | 2026-06-01</p>
</blockquote><h2 id="-论文信息" class="headerLink">
    <a href="#-%e8%ae%ba%e6%96%87%e4%bf%a1%e6%81%af" class="header-mark"></a>📄 论文信息</h2><table>
  <thead>
      <tr>
          <th>项目</th>
          <th>内容</th>
      </tr>
  </thead>
  <tbody>
      <tr>
          <td><strong>标题</strong></td>
          <td>SkySense V2: A Unified Foundation Model for Multi-modal Remote Sensing</td>
      </tr>
      <tr>
          <td><strong>作者</strong></td>
          <td>蚂蚁集团、武汉大学</td>
      </tr>
      <tr>
          <td><strong>会议</strong></td>
          <td>ICCV 2025</td>
      </tr>
      <tr>
          <td><strong>arXiv</strong></td>
          <td>待确认</td>
      </tr>
      <tr>
          <td><strong>GitHub</strong></td>
          <td><a href="https://github.com/kang-wu/SkySensePlusPlus" target="_blank" rel="noopener noreferrer">https://github.com/kang-wu/SkySensePlusPlus</a></td>
      </tr>
      <tr>
          <td><strong>关键词</strong></td>
          <td>遥感基础模型、多模态统一、Transformer、自监督学习、专家混合</td>
      </tr>
  </tbody>
</table>
<h2 id="-解决的核心问题" class="headerLink">
    <a href="#-%e8%a7%a3%e5%86%b3%e7%9a%84%e6%a0%b8%e5%bf%83%e9%97%ae%e9%a2%98" class="header-mark"></a>🎯 解决的核心问题</h2><h3 id="问题背景" class="headerLink">
    <a href="#%e9%97%ae%e9%a2%98%e8%83%8c%e6%99%af" class="header-mark"></a>问题背景</h3><p>在遥感领域，我们经常需要处理多种模态的数据：光学图像（RGB）、多光谱图像（MS）、合成孔径雷达（SAR）等。这些不同模态的数据各有优势——光学图像色彩丰富，SAR能穿透云雾，多光谱能捕捉植被健康状况。</p>
<h3 id="现有方法的局限" class="headerLink">
    <a href="#%e7%8e%b0%e6%9c%89%e6%96%b9%e6%b3%95%e7%9a%84%e5%b1%80%e9%99%90" class="header-mark"></a>现有方法的局限</h3><p>目前的多模态遥感基础模型存在一个尴尬的问题：</p>
<ol>
<li><strong>参数冗余</strong>：为每种模态训练单独的backbone，导致模型参数量爆炸</li>
<li><strong>效率低下</strong>：不同模态的模型无法共享特征，计算资源浪费严重</li>
<li><strong>语义不一致</strong>：不同模态提取的特征难以对齐，影响下游任务性能</li>
</ol>
<h3 id="核心问题提炼" class="headerLink">
    <a href="#%e6%a0%b8%e5%bf%83%e9%97%ae%e9%a2%98%e6%8f%90%e7%82%bc" class="header-mark"></a>核心问题提炼</h3><p><strong>如何用一个统一的backbone高效处理多种遥感模态，同时保持各模态的独特特性？</strong></p>
<h2 id="-解决方案" class="headerLink">
    <a href="#-%e8%a7%a3%e5%86%b3%e6%96%b9%e6%a1%88" class="header-mark"></a>💡 解决方案</h2><h3 id="核心创新点1统一transformer骨干网络" class="headerLink">
    <a href="#%e6%a0%b8%e5%bf%83%e5%88%9b%e6%96%b0%e7%82%b91%e7%bb%9f%e4%b8%80transformer%e9%aa%a8%e5%b9%b2%e7%bd%91%e7%bb%9c" class="header-mark"></a>核心创新点1：统一Transformer骨干网络</h3><p><strong>设计动机</strong>：既然不同模态的图像都是2D数据，为什么不共享一个backbone？</p>
<p><strong>具体实现</strong>：</p>
<div class="code-block highlight is-closed show-line-numbers  tw-group tw-my-2">
  <div class="
    
    tw-flex 
    tw-flex-row
    tw-flex-1 
    tw-justify-between 
    tw-w-full tw-bg-bgColor-secondary
    ">      
    <button 
      class="
        code-block-button
        tw-mx-2 
        tw-flex
        tw-flex-row
        tw-flex-1"
      aria-hidden="true">
          <div class="group-[.is-open]:tw-rotate-90 tw-transition-[transform] tw-duration-500 tw-ease-in-out print:!tw-hidden tw-w-min tw-h-min tw-my-1 tw-mx-1"><svg class="icon"
    xmlns="http://www.w3.org/2000/svg" viewBox="0 0 320 512"><!-- Font Awesome Free 5.15.4 by @fontawesome - https://fontawesome.com License - https://fontawesome.com/license/free (Icons: CC BY 4.0, Fonts: SIL OFL 1.1, Code: MIT License) --><path d="M285.476 272.971L91.132 467.314c-9.373 9.373-24.569 9.373-33.941 0l-22.667-22.667c-9.357-9.357-9.375-24.522-.04-33.901L188.505 256 34.484 101.255c-9.335-9.379-9.317-24.544.04-33.901l22.667-22.667c9.373-9.373 24.569-9.373 33.941 0L285.475 239.03c9.373 9.372 9.373 24.568.001 33.941z"/></svg></div>
          <p class="tw-select-none !tw-my-1">text</p>]]></description></item></channel></rss>