{"id":38099,"date":"2026-03-19T17:30:27","date_gmt":"2026-03-19T09:30:27","guid":{"rendered":"http:\/\/192.168.10.115\/?p=38099"},"modified":"2026-03-19T17:30:27","modified_gmt":"2026-03-19T09:30:27","slug":"2026-03-19-%e5%a4%8d%e6%97%a6%e5%bc%95%e6%9c%9b%e6%9c%80%e6%96%b0%ef%bc%81dynamicvggt%ef%bc%9a%e9%9d%a2%e5%90%91%e8%87%aa%e5%8a%a8%e9%a9%be%e9%a9%b6%e7%9a%84%e7%bb%9f%e4%b8%804d%e5%8a%a8%e6%80%81","status":"publish","type":"post","link":"http:\/\/222.128.65.89:18086\/index.php\/2026\/03\/19\/38099\/","title":{"rendered":"2026-03-19 \u590d\u65e6&\u5f15\u671b\u6700\u65b0\uff01DynamicVGGT\uff1a\u9762\u5411\u81ea\u52a8\u9a7e\u9a76\u7684\u7edf\u4e004D\u52a8\u6001\u573a\u666f\u91cd\u5efa\u6846\u67b6\uff08CVPR’26\uff09"},"content":{"rendered":"\n

\u539f\u6587\u94fe\u63a5\uff1ahttps:\/\/mp.weixin.qq.com\/s\/pSy10dmP9G6BaadYh6OPZw<\/a><\/p>\n\n\n\n

\u8fd1\u5e74\u6765\uff0c3D foundation model \u5728\u51e0\u4f55\u611f\u77e5\u548c\u9759\u6001\u573a\u666f\u91cd\u5efa\u4e0a\u4e0d\u65ad\u7a81\u7834\uff0c\u4f46\u4e00\u5230\u81ea\u52a8\u9a7e\u9a76\u8fd9\u79cd\u771f\u5b9e\u52a8\u6001\u73af\u5883\uff0c\u95ee\u9898\u5c31\u6ca1\u90a3\u4e48\u7b80\u5355\u4e86\u3002\u8f66\u8f86\u5728\u52a8\u3001\u884c\u4eba\u5728\u52a8\u3001\u76f8\u673a\u4e5f\u5728\u52a8\uff0c\u6a21\u578b\u4e0d\u4ec5\u8981\u770b\u6e05\u201c\u573a\u666f\u957f\u4ec0\u4e48\u6837\u201d\uff0c\u8fd8\u8981\u56de\u7b54\u201c\u573a\u666f\u63a5\u4e0b\u6765\u4f1a\u600e\u4e48\u53d8\u201d\u3002\u5982\u4f55\u8ba9 feed-forward 3D \u6a21\u578b\u771f\u6b63\u5177\u5907\u52a8\u6001\u5efa\u6a21\u80fd\u529b\uff0c\u6b63\u5728\u6210\u4e3a 4D \u573a\u666f\u91cd\u5efa\u91cc\u7684\u5173\u952e\u95ee\u9898\u3002<\/p>\n\n\n\n

\u8fd1\u5e74\u6765\uff0c3D foundation model \u5728\u51e0\u4f55\u611f\u77e5\u548c\u9759\u6001\u573a\u666f\u91cd\u5efa\u4e0a\u5c61\u7834\u5929\u82b1\u677f\u2014\u2014\u4f46\u4e00\u65e6\u8fc8\u5165\u81ea\u52a8\u9a7e\u9a76\u7684\u771f\u5b9e\u52a8\u6001\u4e16\u754c\uff0c\u4e8b\u60c5\u5c31\u6ca1\u90a3\u4e48\u8f7b\u677e\u4e86\u3002 \u60f3\u8c61\u4e00\u4e0b\uff1a\u8f66\u8f86\u5728\u7a7f\u884c\uff0c\u884c\u4eba\u5728\u6a2a\u8fc7\u9a6c\u8def\uff0c\u76f8\u673a\u672c\u8eab\u4e5f\u5728\u8fd0\u52a8\u3002\u6a21\u578b\u4e0d\u4ec5\u8981\u56de\u7b54”\u773c\u524d\u7684\u573a\u666f\u957f\u4ec0\u4e48\u6837”\uff0c\u8fd8\u5f97\u7406\u89e3”\u573a\u666f\u63a5\u4e0b\u6765\u4f1a\u600e\u4e48\u53d8”\u3002\u5982\u4f55\u8ba9 feed-forward 3D \u6a21\u578b\u771f\u6b63\u8de8\u5165\u52a8\u6001\u5efa\u6a21\u7684\u95e8\u69db\uff0c\u5df2\u7ecf\u6210\u4e3a 4D \u573a\u666f\u91cd\u5efa\u9886\u57df\u7684\u6838\u5fc3\u6311\u6218\u3002<\/p>\n\n\n\n

\u590d\u65e6\u5927\u5b66<\/strong>\u548c\u5f15\u671b\u667a\u80fd2030Lab<\/strong>\u7684\u7814\u7a76\u56e2\u961f\u63d0\u51fa\uff0c\u628a VGGT \u4ece\u504f\u9759\u6001\u7684 3D \u611f\u77e5\u6846\u67b6\uff0c\u6269\u5c55\u6210\u4e00\u4e2a\u9762\u5411\u81ea\u52a8\u9a7e\u9a76\u573a\u666f\u7684\u7edf\u4e00 4D \u52a8\u6001\u91cd\u5efa\u6846\u67b6\u3002\u5bf9\u6b64\uff0c\u6211\u4eec\u5e76\u6ca1\u6709\u63a8\u5012\u91cd\u6765\uff0c\u800c\u662f\u5728\u73b0\u6709\u51e0\u4f55 backbone \u57fa\u7840\u4e0a\uff0c\u8865\u4e0a\u201c\u65f6\u95f4\u201d\u548c\u201c\u8fd0\u52a8\u201d\u8fd9\u4e24\u5757\u6700\u5173\u952e\u7684\u80fd\u529b\u3002<\/p>\n\n\n\n

\u590d\u65e6\u5927\u5b66\u4e0e\u5f15\u671b\u667a\u80fd2030Lab\u7684\u7814\u7a76\u56e2\u961f\u7ed9\u51fa\u4e86\u4e00\u4e2a\u6e05\u6670\u7684\u56de\u7b54\uff1a\u4e0d\u63a8\u5012\u91cd\u6765\uff0c\u800c\u662f”\u7cbe\u51c6\u8865\u8bfe”\u3002\u4ed6\u4eec\u5728 VGGT \u539f\u6709\u7684\u5f3a\u5927\u51e0\u4f55 backbone \u4e4b\u4e0a\uff0c\u8865\u4e0a\u4e86\u4e24\u5757\u6700\u5173\u952e\u7684\u80fd\u529b\u62fc\u56fe\u2014\u2014”\u65f6\u95f4”\u548c”\u8fd0\u52a8”\uff0c\u7531\u6b64\u6784\u5efa\u51fa\u9762\u5411\u81ea\u52a8\u9a7e\u9a76\u7684\u7edf\u4e00 4D \u52a8\u6001\u91cd\u5efa\u6846\u67b6 DynamicVGGT\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u8bba\u6587\u6807\u9898<\/strong>\uff1aDynamicVGGT: Learning Dynamic Point Maps for 4D Scene Reconstruction in Autonomous Driving<\/p>\n\n\n\n

\u8bba\u6587\u94fe\u63a5<\/strong>\uff1ahttps:\/\/arxiv.org\/abs\/2603.08254<\/p>\n\n\n\n

TL;DR\uff1a<\/strong><\/p>\n\n\n\n

DynamicVGGT \u5229\u7528 VGGT \u539f\u6709\u7a7a\u95f4\u51e0\u4f55\u9884\u6d4b\u80fd\u529b\u4e3a\u57fa\u7840\uff0c\u7528\u65b0\u589e\u7684 MTA blocks \u5b66\u4e60\u52a8\u6001\u573a\u666f\u8fd0\u52a8\uff0c\u6700\u540e\u5206\u522b\u901a\u8fc7\u8bbe\u8ba1\u4e24\u79cd\u4efb\u52a1\uff1a\u672a\u6765\u70b9\u56fe\u9884\u6d4b\u4e0e\u52a8\u6001 4D \u573a\u666f\u91cd\u5efa\u4f18\u5316\u7edf\u4e00\u7684\u52a8\u6001\u70b9\u56fe\uff0c\u589e\u5f3a\u5bf9\u52a8\u6001\u573a\u666f\u4e0b\u6a21\u578b\u5bf9\u51e0\u4f55\u9884\u6d4b\u7684\u80fd\u529b<\/strong><\/p>\n\n\n\n

\u5f53\u524d\u95ee\u9898\uff1a<\/strong><\/p>\n\n\n\n

\u73b0\u6709 feed-forward 3D \u6a21\u578b\u5728\u9759\u6001\u573a\u666f\u51e0\u4f55\u91cd\u5efa\u4e0a\u5df2\u7ecf\u5f88\u5f3a\uff0c\u4f46\u4e00\u5230\u81ea\u52a8\u9a7e\u9a76\u8fd9\u79cd\u771f\u5b9e\u52a8\u6001\u73af\u5883\uff0c\u5c31\u5f88\u96be\u540c\u65f6\u7ef4\u6301\u51e0\u4f55\u7cbe\u5ea6\u548c\u65f6\u95f4\u4e00\u81f4\u6027\u3002\u539f\u56e0\u6709\u4e24\u5c42\u3002\u4e00\u5c42\u662f\u573a\u666f\u672c\u8eab\u52a8\u6001\u590d\u6742\uff0c\u8f66\u3001\u4eba\u3001\u76f8\u673a\u90fd\u5728\u52a8\uff0c\u8fd8\u4f34\u968f\u957f\u65f6\u95f4\u4f9d\u8d56\uff1b\u53e6\u4e00\u5c42\u662f\u81ea\u52a8\u9a7e\u9a76\u6570\u636e\u672c\u8eab\u5f80\u5f80\u5927\u89c4\u6a21\u3001\u9ad8\u566a\u58f0\u3001\u6df1\u5ea6\u7a00\u758f\uff0c\u76f4\u63a5\u8bad\u7ec3\u5f88\u5bb9\u6613\u628a\u539f\u672c\u5728\u9759\u6001\u6570\u636e\u4e0a\u5b66\u5230\u7684\u7a20\u5bc6\u51e0\u4f55\u80fd\u529b\u62c9\u57ae\u3002\u4e3a\u6b64\u6211\u4eec\u63d0\u51fa\uff1a\u505a\u4e00\u4e2a\u7edf\u4e00\u7684\u3001\u524d\u9988\u5f0f\u7684 4D \u573a\u666f\u91cd\u5efa\u6846\u67b6\uff0c\u5728\u4e0d\u4f9d\u8d56\u663e\u5f0f\u76f8\u673a\u5916\u53c2\u5bf9\u9f50\u7684\u524d\u63d0\u4e0b\uff0c\u540c\u65f6\u5efa\u6a21\u51e0\u4f55\u548c\u8fd0\u52a8\u3002<\/p>\n\n\n\n

\u4ece\u5b9a\u4f4d\u4e0a\u8bb2\uff0cDynamicVGGT \u4e0d\u662f\u6539\u53d8\u539f\u6709\u7684VGGT\u8303\u5f0f\uff0c\u800c\u662f\u5efa\u7acb\u5728 VGGT \u4e0a\uff0c\u628a\u5b83\u4ece\u201c\u9759\u6001 3D \u611f\u77e5\u201d\u6269\u5c55\u6210\u201c\u52a8\u6001 4D \u91cd\u5efa\u201d\u3002\u6211\u4eec\u8ba4\u4e3a\uff0cVGGT \u539f\u672c\u7684\u51e0\u4f55 backbone \u5df2\u7ecf\u5f88\u5f3a\uff0c\u95ee\u9898\u4e0d\u5728\u4e8e\u7f3a\u51e0\u4f55\u80fd\u529b\uff0c\u800c\u5728\u4e8e\u7f3a\u4e00\u4e2a\u7edf\u4e00\u7684\u52a8\u6001\u8868\u793a\u548c\u914d\u5957\u7684\u65f6\u5e8f\u5efa\u6a21\u673a\u5236\u3002\u6240\u4ee5\u6574\u7bc7\u8bba\u6587\u7684\u4e3b\u7ebf\u5176\u5b9e\u53ea\u6709\u4e00\u53e5\u8bdd\uff1a\u4ee5 Dynamic Point Map \u4e3a\u7edf\u4e00\u8868\u793a\uff0c\u5728 backbone \u4e2d\u8865\u4e0a\u65f6\u5e8f\u63a8\u7406\uff0c\u5728\u8f93\u51fa\u7aef\u8865\u4e0a\u52a8\u6001\u51e0\u4f55\u7ec6\u5316\u3002<\/p>\n\n\n\n

\u76f4\u9762\u95ee\u9898\uff1a\u9759\u6001\u5f3a\u4e0d\u7b49\u4e8e\u52a8\u6001\u5f3a<\/strong><\/p>\n\n\n\n

\u73b0\u6709\u7684 feed-forward 3D \u6a21\u578b\u5728\u9759\u6001\u573a\u666f\u4e0a\u5df2\u7ecf\u8db3\u591f\u51fa\u8272\uff0c\u4f46\u6362\u5230\u81ea\u52a8\u9a7e\u9a76\u573a\u666f\uff0c\u51e0\u4f55\u7cbe\u5ea6\u548c\u65f6\u95f4\u4e00\u81f4\u6027\u5f80\u5f80\u96be\u4ee5\u517c\u5f97\u3002\u539f\u56e0\u6765\u81ea\u4e24\u4e2a\u5c42\u9762\uff1a<\/p>\n\n\n\n

    \n
  • \u7b2c\u4e00\u5c42\u662f\u573a\u666f\u672c\u8eab\u7684\u590d\u6742\u6027\u3002 \u8f66\u8f86\u3001\u884c\u4eba\u3001\u76f8\u673a\u4e09\u8005\u540c\u65f6\u5728\u8fd0\u52a8\uff0c\u53e0\u52a0\u957f\u8ddd\u79bb\u7684\u65f6\u95f4\u4f9d\u8d56\uff0c\u4ec5\u9760\u5355\u5e27\u51e0\u4f55\u63a8\u7406\u8fdc\u8fdc\u4e0d\u591f\u3002<\/li>\n\n\n\n
  • \u7b2c\u4e8c\u5c42\u662f\u6570\u636e\u5c42\u9762\u7684\u5148\u5929\u7f3a\u9677\u3002 \u81ea\u52a8\u9a7e\u9a76\u6570\u636e\u89c4\u6a21\u5927\u3001\u566a\u58f0\u9ad8\u3001\u6df1\u5ea6\u6807\u6ce8\u7a00\u758f\u2014\u2014\u76f4\u63a5\u5728\u8fd9\u4e9b\u6570\u636e\u4e0a\u8bad\u7ec3\uff0c\u6781\u6613”\u62d6\u57ae”\u6a21\u578b\u5728\u9759\u6001\u6570\u636e\u4e0a\u8f9b\u82e6\u5b66\u6765\u7684\u7a20\u5bc6\u51e0\u4f55\u80fd\u529b\u3002<\/li>\n<\/ul>\n\n\n\n

    \u9762\u5bf9\u8fd9\u4e24\u4e2a\u74f6\u9888\uff0cDynamicVGGT \u7684\u5b9a\u4f4d\u975e\u5e38\u660e\u786e\uff1a\u4e0d\u662f\u66ff\u6362 VGGT\uff0c\u800c\u662f\u7ad9\u5728\u5b83\u80a9\u8180\u4e0a\u2014\u2014\u628a”\u9759\u6001 3D \u611f\u77e5”\u5e73\u6ed1\u5730\u6269\u5c55\u4e3a”\u52a8\u6001 4D \u91cd\u5efa”\uff0c\u540c\u65f6\u5728\u4e0d\u4f9d\u8d56\u663e\u5f0f\u76f8\u673a\u5916\u53c2\u5bf9\u9f50\u7684\u524d\u63d0\u4e0b\uff0c\u7edf\u4e00\u5efa\u6a21\u51e0\u4f55\u4e0e\u8fd0\u52a8\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u65b9\u6cd5<\/strong><\/p>\n\n\n\n

    DynamicVGGT framework<\/strong><\/p>\n\n\n\n

    \u4eec\u7ed9\u5b9a\u4e00\u4e2a\u591a\u89c6\u89d2 clip\uff0c\u76f4\u63a5\u540c\u65f6\u9884\u6d4b\u5f53\u524d\u65f6\u523b\u548c\u672a\u6765\u65f6\u523b\u7684 point map\uff0c\u5e76\u4e14\u4e8c\u8005\u90fd\u5728\u8fd9\u4e2a\u5171\u4eab\u5b66\u4e60\u5750\u6807\u7a7a\u95f4\u91cc\u8868\u8fbe\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u6a21\u578b\u5c31\u80fd\u901a\u8fc7\u5f53\u524d\u70b9\u56fe\u548c\u672a\u6765\u70b9\u56fe\u4e4b\u95f4\u7684\u5dee\u5206\uff0c\u9690\u5f0f\u5b66\u5230\u70b9\u7684\u8fd0\u52a8\u3002<\/p>\n\n\n\n

    \u5728\u8fd9\u4e2a\u7edf\u4e00\u8868\u793a\u4e0a\uff0c\u8bba\u6587\u8bbe\u8ba1\u4e86\u4e24\u7c7b\u4e92\u8865\u4efb\u52a1\u3002\u7b2c\u4e00\u7c7b\u662f Future Point Head\uff0c\u5bf9\u672a\u6765\u70b9\u56fe\u8fdb\u884c\u9884\u6d4b\uff1b\u7b2c\u4e8c\u7c7b\u662f Dynamic 3D Gaussian Splatting Head\uff0c\u5bf9\u52a8\u6001\u51e0\u4f55\u505a\u66f4\u663e\u5f0f\u7684\u7ec6\u5316\u3002\u6211\u4eec\u628a\u8fd9\u4e24\u8005\u660e\u786e\u770b\u6210\u4e92\u8865\u5173\u7cfb\uff1a\u524d\u8005\u504f\u201c\u9690\u5f0f\u8fd0\u52a8\u5b66\u4e60\u201d\uff0c\u540e\u8005\u504f\u201c\u663e\u5f0f\u8fd0\u52a8\u76d1\u7763\u201d\u3002\u8fd9\u4e5f\u662f\u8bba\u6587\u65b9\u6cd5\u8bbe\u8ba1\u4e2d\u6700\u6e05\u695a\u7684\u4e00\u70b9\uff0c\u4e0d\u662f\u53ea\u7528\u4e00\u79cd\u52a8\u6001\u4fe1\u53f7\uff0c\u800c\u662f\u540c\u65f6\u4ece point-map \u5c42\u548c Gaussian primitive \u5c42\u53bb\u7ea6\u675f\u52a8\u6001\u3002<\/p>\n\n\n\n

    \u8fd9\u5f20\u56fe\u5c31\u662f DynamicVGGT \u7684\u6574\u4f53\u8bad\u7ec3\u6846\u67b6\u56fe\u3002\u53ef\u4ee5\u628a\u5b83\u7406\u89e3\u6210\u4e00\u6761\u5f88\u6e05\u6670\u7684\u6d41\u6c34\u7ebf\uff1a<\/p>\n\n\n\n

    \u8f93\u5165\u591a\u89c6\u89d2\u56fe\u50cf\u5e8f\u5217 \u2192 \u63d0\u53d6\u7a7a\u95f4\u51e0\u4f55\u7279\u5f81 \u2192 \u5efa\u6a21\u65f6\u5e8f\u8fd0\u52a8 \u2192 \u5206\u4e24\u6761\u5934\u5206\u522b\u505a\u672a\u6765\u70b9\u56fe\u9884\u6d4b\u548c\u52a8\u6001\u9ad8\u65af\u91cd\u5efa\u3002<\/strong><\/p>\n\n\n\n

    \u6700\u5de6\u8fb9\u8f93\u5165\u7684\u662f\u4e00\u4e2a\u591a\u89c6\u89d2\u56fe\u50cf\u5e8f\u5217\uff0c\u8fd9\u91cc\u753b\u6210\u4e86 {V1, V2, V3}<\/code>\u3002\u6bcf\u4e2a Vi<\/code> \u53ef\u4ee5\u7406\u89e3\u4e3a\u540c\u4e00\u65f6\u523b\u6216\u76f8\u90bb\u65f6\u523b\u7684\u4e0d\u540c\u76f8\u673a\u89c6\u89d2\u56fe\u50cf\u3002<\/p>\n\n\n\n

    \u7b2c\u4e00\u6b65\u662f DINOv2 Backbone<\/strong>\u3002\u8fd9\u90e8\u5206\u8d1f\u8d23\u628a\u539f\u59cb\u56fe\u50cf\u7f16\u7801\u6210 token \u8868\u793a\u3002\u8f93\u51fa\u4e24\u7c7b token\uff1a\u4e00\u7c7b\u662f patch tokens<\/strong>\uff0c\u8868\u793a\u56fe\u50cf\u5c40\u90e8\u533a\u57df\u7684\u89c6\u89c9\u7279\u5f81\uff1b\u53e6\u4e00\u7c7b\u662f camera tokens<\/strong>\uff0c\u8868\u793a\u548c\u76f8\u673a\u76f8\u5173\u7684\u5168\u5c40\u4fe1\u606f\u3002<\/p>\n\n\n\n

    \u63a5\u7740\u8fdb\u5165\u4e2d\u95f4\u7684 Alternating-Attention (AA) blocks<\/strong>\u3002\u8fd9\u90e8\u5206\u7ee7\u627f\u81ea VGGT\uff0c\u672c\u8d28\u4e0a\u8fd8\u662f\u539f\u59cb backbone \u91cc\u6700\u91cd\u8981\u7684\u7a7a\u95f4\u51e0\u4f55\u5efa\u6a21\u6a21\u5757\u3002\u5b83\u7684\u4f5c\u7528\u662f\u5bf9 patch tokens \u548c camera tokens \u505a\u5e27\u5185\u5efa\u6a21\uff0c\u8ba9\u6a21\u578b\u5148\u5b66\u4f1a\u201c\u8fd9\u4e00\u5e27\u91cc\u7a7a\u95f4\u7ed3\u6784\u662f\u4ec0\u4e48\u6837\u7684\u201d\u3002\u6240\u4ee5 AA blocks \u4e3b\u8981\u8d1f\u8d23\u7684\u662f intra-frame spatial geometry<\/strong>\uff0c\u4e5f\u5c31\u662f\u5355\u5e27\u5185\u90e8\u7684\u51e0\u4f55\u5173\u7cfb\uff0c\u800c\u4e0d\u662f\u65f6\u95f4\u4e0a\u7684\u8fd0\u52a8\u5173\u7cfb\u3002<\/p>\n\n\n\n

    \u548c AA blocks \u5e76\u884c\u7684\uff0c\u5c31\u662f\u56fe\u4e2d\u4e0b\u9762\u8fd9\u6761 Motion-aware Temporal Attention (MTA) blocks<\/strong>\u3002\u8fd9\u662f DynamicVGGT \u6700\u5173\u952e\u7684\u65b0\u52a0\u6a21\u5757\u4e4b\u4e00\u3002\u5b83\u7684\u8f93\u5165\u9664\u4e86\u6765\u81ea\u4e0a\u9762\u7a7a\u95f4\u5206\u652f\u7684 patch tokens\uff0c\u8fd8\u989d\u5916\u5f15\u5165\u4e86\u56fe\u4e2d\u84dd\u8272\u7684\u5c0f\u5757 motion tokens<\/strong>\u3002\u8fd9\u4e9b motion tokens \u662f\u53ef\u5b66\u4e60\u53c2\u6570\uff0c\u4e13\u95e8\u7528\u6765\u7f16\u7801\u65f6\u95f4\u5148\u9a8c\u548c\u8de8\u5e27\u8fd0\u52a8\u7ebf\u7d22\u3002MTA \u7684\u4f5c\u7528\uff0c\u5c31\u662f\u6cbf\u65f6\u95f4\u7ef4\u53bb\u770b\u4e0d\u540c\u5e27\u4e4b\u95f4\u7684\u5173\u7cfb\uff0c\u5efa\u6a21 inter-frame temporal dependencies<\/strong>\u3002\u6362\u53e5\u8bdd\u8bf4\uff0cAA \u8d1f\u8d23\u56de\u7b54\u201c\u6bcf\u5e27\u957f\u4ec0\u4e48\u6837\u201d\uff0cMTA \u8d1f\u8d23\u56de\u7b54\u201c\u76f8\u90bb\u5e27\u4e4b\u95f4\u600e\u4e48\u53d8\u5316\u201d\u3002<\/p>\n\n\n\n

    \u7ecf\u8fc7 MTA \u4e4b\u540e\uff0c\u6a21\u578b\u5f97\u5230\u7684\u662f\u56fe\u4e2d\u53f3\u8fb9\u6807\u51fa\u6765\u7684 Temporal Features TA<\/strong>\u3002\u8fd9\u90e8\u5206\u5c31\u662f\u201c\u5df2\u7ecf\u878d\u5408\u4e86\u65f6\u95f4\u4fe1\u606f\u201d\u7684\u7279\u5f81\uff0c\u53ef\u4ee5\u7406\u89e3\u6210\u52a8\u6001\u589e\u5f3a\u540e\u7684\u8868\u793a\u3002\u8bba\u6587\u540e\u9762\u51e0\u4e4e\u6240\u6709\u52a8\u6001\u4efb\u52a1\uff0c\u90fd\u662f\u57fa\u4e8e\u8fd9\u4e2a TA<\/code> \u6765\u505a\u7684\u3002<\/p>\n\n\n\n

    \u7136\u540e\u56fe\u53f3\u8fb9\u5206\u6210\u4e24\u6761\u4efb\u52a1\u5934\u3002<\/p>\n\n\n\n

    \u4e0a\u9762\u8fd9\u4e00\u6761\u662f DGSHead<\/strong>\uff0c\u4e5f\u5c31\u662f Dynamic 3D Gaussian Head<\/strong>\u3002\u5b83\u8d1f\u8d23\u505a\u52a8\u6001 3D Gaussian Splatting \u91cd\u5efa\u3002\u4f60\u53ef\u4ee5\u628a\u5b83\u7406\u89e3\u6210\uff1a\u6a21\u578b\u4e0d\u53ea\u662f\u6062\u590d\u70b9\u56fe\uff0c\u8fd8\u8981\u628a\u573a\u666f\u8868\u793a\u6210\u53ef\u968f\u65f6\u95f4\u53d8\u5316\u7684 3D Gaussian primitives\uff0c\u5e76\u4e14\u663e\u5f0f\u9884\u6d4b\u5b83\u4eec\u7684\u8fd0\u52a8\u3002\u8fd9\u4e2a head \u66f4\u504f\u5411\u663e\u5f0f\u52a8\u6001\u51e0\u4f55\u5efa\u6a21<\/strong>\uff0c\u6700\u540e\u670d\u52a1\u4e8e\u52a8\u6001\u573a\u666f\u91cd\u5efa\u548c novel view synthesis\u3002<\/p>\n\n\n\n

    \u4e0b\u9762\u8fd9\u4e00\u6761\u662f FPH\uff08Future Point Head\uff09<\/strong>\u3002\u5b83\u8d1f\u8d23\u4ece\u5f53\u524d\u65f6\u523b\u7684 temporal feature \u76f4\u63a5\u9884\u6d4b\u672a\u6765\u65f6\u523b\u7684 point map\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u5b83\u8ba9\u6a21\u578b\u5b66\u4e60\u201c\u4e0b\u4e00\u5e27\u7684\u4e09\u7ef4\u70b9\u56fe\u5e94\u8be5\u957f\u4ec0\u4e48\u6837\u201d\u3002\u8fd9\u4e2a\u4efb\u52a1\u672c\u8d28\u4e0a\u662f\u5728\u505a future point prediction<\/strong>\uff0c\u901a\u8fc7\u9884\u6d4b\u672a\u6765\u70b9\u56fe\uff0c\u8ba9\u6a21\u578b\u9690\u5f0f\u5b66\u4f1a\u70b9\u7ea7\u8fd0\u52a8\u3002<\/p>\n\n\n\n

    \u6240\u4ee5\uff0c\u8fd9\u4e24\u6761\u5934\u867d\u7136\u90fd\u548c\u52a8\u6001\u6709\u5173\uff0c\u4f46\u5206\u5de5\u4e0d\u4e00\u6837\uff1a<\/p>\n\n\n\n

      \n
    • DGSHead<\/strong>\u00a0\u66f4\u504f\uff1a\u663e\u5f0f\u8fd0\u52a8\u5efa\u6a21 + \u52a8\u6001\u9ad8\u65af\u91cd\u5efa<\/li>\n\n\n\n
    • FPH<\/strong>\u00a0\u66f4\u504f\uff1a\u9690\u5f0f\u8fd0\u52a8\u5b66\u4e60 + \u672a\u6765\u51e0\u4f55\u9884\u6d4b<\/li>\n<\/ul>\n\n\n\n

      DynamicVGGT \u7684\u89e3\u51b3\u601d\u8def\u662f\uff1a\u8ba9\u6a21\u578b\u540c\u65f6\u9884\u6d4b\u5f53\u524d\u5e27\u548c\u672a\u6765\u5e27\u7684\u70b9\u56fe\uff0c\u5e76\u5728\u540c\u4e00\u53c2\u8003\u5750\u6807\u7cfb\u4e0b\u5bf9\u9f50\u3002\u8fd9\u6837\uff0c\u6a21\u578b\u53ef\u4ee5\u901a\u8fc7\u65f6\u5e8f\u5bf9\u5e94\u5173\u7cfb\uff0c\u9690\u5f0f\u5730\u5b66\u4f1a”\u7269\u4f53\u662f\u600e\u4e48\u52a8\u7684”\uff0c\u800c\u65e0\u9700\u4efb\u4f55\u663e\u5f0f\u7684\u8fd0\u52a8\u6807\u6ce8\u3002<\/p>\n\n\n\n

      \u66f4\u5173\u952e\u7684\u662f\uff0c\u6574\u4e2a\u8fc7\u7a0b\u65e0\u9700\u76f8\u673a\u5916\u53c2\u5bf9\u9f50\uff0c\u6781\u5927\u964d\u4f4e\u4e86\u771f\u5b9e\u90e8\u7f72\u7684\u95e8\u69db\u3002<\/p>\n\n\n\n

      \u4e09\u5927\u6838\u5fc3\u6a21\u5757\uff1a\u4ece”\u611f\u77e5\u5f53\u4e0b”\u5230”\u7406\u89e3\u65f6\u5e8f”<\/p>\n\n\n\n

      \u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0c\u8bba\u6587\u8bbe\u8ba1\u4e86\u4e09\u4e2a\u7d27\u5bc6\u534f\u4f5c\u7684\u5173\u952e\u6a21\u5757\uff1a<\/p>\n\n\n\n

      \u8fd0\u52a8\u611f\u77e5\u65f6\u5e8f\u6ce8\u610f\u529b\uff08MTA\uff09\uff1a\u5728\u6807\u51c6\u7684\u591a\u89c6\u89d2 Transformer \u4e2d\u5f15\u5165\u65f6\u5e8f\u7ef4\u5ea6\uff0c\u8ba9\u6a21\u578b\u4e0d\u53ea”\u770b\u5f53\u524d\u5e27”\uff0c\u800c\u662f\u5b66\u4f1a\u8de8\u5e27\u7684\u8fd0\u52a8\u8fde\u7eed\u6027\u3002\u5c31\u50cf\u4eba\u7c7b\u770b\u89c6\u9891\u65f6\u4f1a\u81ea\u7136\u5730\u8ffd\u8e2a\u8fd0\u52a8\u7269\u4f53\uff0c\u800c\u975e\u6bcf\u5e27\u72ec\u7acb\u7406\u89e3\u3002<\/p>\n\n\n\n

      \u672a\u6765\u70b9\u56fe\u9884\u6d4b\u5934\uff08Future Point Head\uff09\uff1a\u5728\u8054\u5408\u8bad\u7ec3\u7684\u6846\u67b6\u4e0b\uff0c\u9884\u6d4b\u573a\u666f\u4e2d\u6bcf\u4e2a\u70b9\u5728\u672a\u6765\u65f6\u523b\u7684\u4f4d\u7f6e\uff0c\u4e3a\u540e\u7eed\u7684\u52a8\u6001\u9ad8\u65af\u4f18\u5316\u63d0\u4f9b\u8fd0\u52a8\u5148\u9a8c\u3002<\/p>\n\n\n\n

      \u52a8\u60013D\u9ad8\u65af\u6cfc\u6e85\u6e32\u67d3\u5934\uff08Dynamic 3D Gaussian Splatting Head\uff09\uff1a\u5728\u4e0a\u8ff0\u8fd0\u52a8\u8868\u5f81\u7684\u57fa\u7840\u4e0a\uff0c\u5f15\u5165\u53ef\u5b66\u4e60\u7684\u8fd0\u52a8 token\uff0c\u901a\u8fc7\u573a\u666f\u6d41\u76d1\u7763\u6765\u9884\u6d4b\u6bcf\u4e2a\u9ad8\u65af\u7403\u7684\u8fd0\u52a8\u901f\u5ea6\uff0c\u5e76\u4ee5\u6b64\u6301\u7eed\u4f18\u5316\u51e0\u4f55\u8d28\u91cf\u3002\u6700\u7ec8\u5b9e\u73b0\u9ad8\u8d28\u91cf\u7684\u65b0\u89c6\u89d2\u5408\u6210\uff0c\u54ea\u6015\u662f\u5728\u6709\u5927\u91cf\u8fd0\u52a8\u7269\u4f53\u7684\u9a7e\u9a76\u573a\u666f\u4e2d\u3002<\/p>\n\n\n\n

      \u5b9e\u9a8c\u7ed3\u679c<\/strong><\/p>\n\n\n\n

      \"\"<\/figure>\n\n\n\n
      \"\"<\/figure>\n\n\n\n

      \u8bba\u6587\u5728\u4e24\u5927\u6743\u5a01\u81ea\u52a8\u9a7e\u9a76\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u4e86\u5168\u9762\u8bc4\u6d4b\uff1a<\/p>\n\n\n\n

      \u573a\u666f\u91cd\u5efa\u7cbe\u5ea6\uff1a\u5728 Accuracy\u3001Completeness\u3001Normal Consistency \u7b49\u6838\u5fc3\u6307\u6807\u4e0a\uff0cDynamicVGGT \u663e\u8457\u4f18\u4e8e\u6b64\u524d\u6240\u6709\u5bf9\u6bd4\u65b9\u6cd5\uff1b<\/p>\n\n\n\n

      \u65b0\u89c6\u89d2\u5408\u6210\uff08NVS\uff09\uff1a\u7ed9\u5b9a\u7a00\u758f\u8f93\u5165\u5e27\uff08\u5982\u7b2c0\u30012\u30014\u5e27\uff09\uff0c\u6a21\u578b\u4e0d\u4ec5\u80fd\u91cd\u5efa\u5b8c\u6574\u573a\u666f\uff0c\u8fd8\u80fd\u76f4\u63a5\u5408\u6210\u7b2c5\u5e27\u7684\u771f\u5b9e\u611f\u6e32\u67d3\u7ed3\u679c\uff0c\u65f6\u5e8f\u4e00\u81f4\u6027\u51fa\u8272\uff1b<\/p>\n\n\n\n

      \u591a\u4efb\u52a1\u8f93\u51fa\u4e30\u5bcc\uff1a\u9664\u4e86\u573a\u666f\u91cd\u5efa\uff0c\u6a21\u578b\u540c\u65f6\u8f93\u51fa\u53ef\u7528\u7684\u6df1\u5ea6\u56fe\u3001\u76f8\u673a\u4f4d\u59ff\u4f30\u8ba1\uff0c\u4e00\u4e2a\u524d\u9988\u7f51\u7edc\u641e\u5b9a\u591a\u4e2a\u81ea\u52a8\u9a7e\u9a76\u611f\u77e5\u5b50\u4efb\u52a1\u3002<\/p>\n\n\n\n

      DynamicVGGT \u5728 point map reconstruction \u4e0a\u6bd4 VGGT \u548c StreamVGGT \u66f4\u597d\u3002KITTI \u5355\u76ee\u4e09\u5e27\u8bbe\u7f6e\u4e0b\uff0c\u5b83\u8fbe\u5230 0.901 \u7684 Accuracy<\/strong> \u548c 0.939 \u7684 Normal Consistency<\/strong>\uff1bWaymo \u4e09\u76f8\u673a\u8bbe\u7f6e\u4e0b\uff0c\u5b83\u8fbe\u5230 4.021 \u7684 Accuracy<\/strong> \u548c 0.603 \u7684 Normal Consistency<\/strong>\u3002<\/p>\n\n\n\n

      \u8fd9\u4e2a\u6a21\u578b\u5728\u6df1\u5ea6\u4f30\u8ba1\u4e0a\u4e5f\u8868\u73b0\u5f88\u5f3a\uff0c\u8bf4\u660e\u52a8\u6001\u5efa\u6a21\u5e76\u6ca1\u6709\u7834\u574f\u51e0\u4f55 backbone\uff0c\u53cd\u800c\u589e\u5f3a\u4e86\u5b83\u7684\u6cdb\u5316\u80fd\u529b\u3002KITTI \u5355\u76ee Abs Rel 0.070<\/strong>\uff0cNYU-v2 \u5355\u76ee Abs Rel 0.064<\/strong>\u3001\u03b4 < 1.25 \u51c6\u786e\u7387 0.943<\/strong>\uff0cKITTI MVS Abs Rel 0.051<\/strong>\u3001\u51c6\u786e\u7387 0.976<\/strong>\uff0c\u6574\u4f53\u4f18\u4e8e VGGT \u548c StreamVGGT\u3002\u8fd9\u4e2a\u7ed3\u679c\u5176\u5b9e\u633a\u8bf4\u660e\u95ee\u9898\uff1aDynamicVGGT \u4e0d\u662f\u4e3a\u4e86\u52a8\u6001\u4efb\u52a1\u727a\u7272\u51e0\u4f55\uff0c\u800c\u662f\u5728\u66f4\u5f3a\u65f6\u95f4\u5efa\u6a21\u4e0b\u628a\u51e0\u4f55\u4e5f\u4e00\u8d77\u505a\u7a33\u4e86\u3002<\/p>\n\n\n\n

      \u6d88\u878d\u5b9e\u9a8c<\/strong><\/p>\n\n\n\n

      \u6211\u4eec\u4ee5 vanilla VGGT \u4e3a\u57fa\u7ebf\uff0c\u5148\u52a0 temporal attention \u548c FPH\uff0c\u5728 KITTI \u4e0a Accuracy \u4ece\u00a01.489<\/strong>\u00a0\u964d\u5230\u00a00.927<\/strong>\uff0cCompleteness \u4ece\u00a00.690<\/strong>\u00a0\u964d\u5230\u00a00.600<\/strong>\uff1b\u518d\u52a0 DGSHead\uff0cAccuracy \u8fdb\u4e00\u6b65\u5230\u00a00.901<\/strong>\uff0cNormal Consistency \u5230\u00a00.939<\/strong>\u3002Waymo \u4e0a\u5b8c\u6574\u6a21\u578b\u4e5f\u8fbe\u5230\u6700\u4f18\u7684\u00a04.021 Accuracy error<\/strong>\u00a0\u548c\u00a00.603 NC<\/strong>\u3002\u8fd9\u4e2a\u6d88\u878d\u8bf4\u660e\u4e24\u4ef6\u4e8b\uff1a\u4e00\u662f\u201c\u65f6\u95f4\u5efa\u6a21\u201d\u672c\u8eab\u5df2\u7ecf\u5e26\u6765\u5f88\u5927\u6536\u76ca\uff1b\u4e8c\u662f\u201c\u52a8\u6001 Gaussian \u7ec6\u5316\u201d\u4f1a\u8fdb\u4e00\u6b65\u628a\u51e0\u4f55\u53d8\u5f97\u66f4\u5e73\u6ed1\u3001\u66f4\u5b8c\u6574\u3001\u66f4\u4e00\u81f4\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u8fd9\u4e0d\u662f\u5355\u4e2a\u6a21\u5757\u5076\u7136\u594f\u6548\uff0c\u800c\u662f\u6574\u6761\u00a0DPM + MTA + FPH + DGSHead<\/strong>\u00a0\u7684\u94fe\u6761\u90fd\u5728\u8d77\u4f5c\u7528\u3002<\/p>\n\n\n\n

      \"\"<\/figure>\n\n\n\n

      \u53ef\u89c6\u5316<\/strong><\/p>\n\n\n\n

      \u53ef\u89c6\u5316\u7ed3\u679c\u8fdb\u4e00\u6b65\u652f\u6491\u4e86\u8fd9\u4e2a\u7ed3\u8bba\u3002\u8bba\u6587\u5c55\u793a\u4e86\u5355\u5e27\u3001\u77ed\u65f6\u591a\u5e27\u3001\u957f\u65f6\u5e8f\u4e09\u79cd\u8bbe\u7f6e\u4e0b\u7684\u70b9\u56fe\u91cd\u5efa\uff0c\u7ed3\u8bba\u90fd\u662f DynamicVGGT \u6bd4 VGGT \u66f4\u7a20\u5bc6\u3001\u66f4\u5e73\u6ed1\uff0c\u800c\u4e14\u5728\u5927\u89c6\u89d2\u53d8\u5316\u3001\u4e0b\u5761\u8def\u3001\u5f00\u9614\u8def\u53e3\u8fd9\u4e9b\u96be\u573a\u666f\u4e0b\uff0c\u4ecd\u7136\u80fd\u4fdd\u6301\u7a33\u5b9a\u7684\u5168\u5c40\u51e0\u4f55\u4e0e\u70b9\u7ea7\u8fd0\u52a8\u8f68\u8ff9\u3002<\/p>\n\n\n\n

      \"\"<\/figure>\n\n\n\n

      \u603b\u7ed3<\/strong><\/p>\n\n\n\n

      DynamicVGGT \u7684\u6838\u5fc3\u4e0d\u662f\u7b80\u5355\u7ed9 VGGT \u52a0\u65f6\u5e8f\u6a21\u5757\uff0c\u800c\u662f\u56f4\u7ed5\u7edf\u4e00\u7684 Dynamic Point Map<\/strong> \u8868\u793a\uff0c\u628a\u52a8\u6001\u5efa\u6a21\u62c6\u6210\u4e09\u5c42\uff1a<\/p>\n\n\n\n

        \n
      • MTA<\/strong>\u00a0\u5728\u7279\u5f81\u5c42\u5efa\u6a21\u65f6\u95f4\u4f9d\u8d56<\/li>\n\n\n\n
      • FPH<\/strong>\u00a0\u5728\u70b9\u56fe\u5c42\u9690\u5f0f\u5b66\u4e60\u8de8\u5e27\u8fd0\u52a8<\/li>\n\n\n\n
      • DGSHead<\/strong>\u00a0\u5728 Gaussian primitive \u5c42\u663e\u5f0f\u5b66\u4e60\u901f\u5ea6\u5e76\u7ec6\u5316\u52a8\u6001\u51e0\u4f55<\/li>\n<\/ul>\n\n\n\n

        \u518d\u901a\u8fc7\u5408\u6210\u5230\u771f\u5b9e\u7684\u4e24\u9636\u6bb5\u8bad\u7ec3\uff0c\u517c\u987e\u51e0\u4f55\u5148\u9a8c\u3001\u8fd0\u52a8\u8fde\u7eed\u6027\u548c\u771f\u5b9e\u9a7e\u9a76\u573a\u666f\u9002\u5e94\u6027\uff0c\u6700\u7ec8\u628a\u9759\u6001 3D \u611f\u77e5\u63a8\u8fdb\u6210\u7edf\u4e00\u7684 feed-forward 4D \u52a8\u6001\u573a\u666f\u91cd\u5efa\u6846\u67b6<\/strong>\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

        \u539f\u6587\u94fe\u63a5\uff1ahttps:\/\/mp.weixin. […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3,1],"tags":[],"class_list":["post-38099","post","type-post","status-publish","format-standard","hentry","category-technology-frontier","category-home"],"views":0,"_links":{"self":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/38099"}],"collection":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/comments?post=38099"}],"version-history":[{"count":1,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/38099\/revisions"}],"predecessor-version":[{"id":38106,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/38099\/revisions\/38106"}],"wp:attachment":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/media?parent=38099"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/categories?post=38099"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/tags?post=38099"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}