{"id":789,"date":"2024-07-15T17:27:13","date_gmt":"2024-07-15T09:27:13","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=789"},"modified":"2024-07-25T15:16:12","modified_gmt":"2024-07-25T07:16:12","slug":"202407-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-22%e7%ba%a7%e9%bb%8e%e9%92%b0-pyramid-vision-transformera-versatile-backbone-for-dense-prediction-without-convolutions","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202407-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-22%e7%ba%a7%e9%bb%8e%e9%92%b0-pyramid-vision-transformera-versatile-backbone-for-dense-prediction-without-convolutions\/","title":{"rendered":"202407 \u8bba\u6587\u7814\u8bfb-Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions"},"content":{"rendered":"\n

\u6765\u6e90\uff1a2021 IEEE\/CVF International Conference on Computer Vision (ICCV)<\/p>\n\n\n\n

\u4f5c\u8005\uff1aWenhai Wang, Enze Xie, Xiang Li\u7b49<\/p>\n\n\n\n

\u5355\u4f4d\uff1aNanjing University of Science and Technology<\/p>\n\n\n\n

\u4e00\u3001\u5f15\u8a00<\/h2>\n\n\n\n

\u4f20\u7edf\u7684ViT\u4e3a\u67f1\u72b6\u7ed3\u6784\uff0c\u8f93\u5165\u7c97\u56fe\u50cf\u5757\u3002\u867d\u7136ViT\u9002\u7528\u4e8e\u56fe\u50cf\u5206\u7c7b\uff0c\u4f46\u5c06\u5176\u76f4\u63a5\u5e94\u7528\u4e8e\u76ee\u6807\u68c0\u6d4b\u548c\u5206\u5272\u7b49\u50cf\u7d20\u7ea7\u5bc6\u96c6\u9884\u6d4b\u5177\u6709\u6311\u6218\u6027\uff0c\u56e0\u4e3a\uff1a<\/p>\n\n\n\n

(1)ViT\u7684\u8f93\u51fa\u5e8f\u5217\u7684\u957f\u5ea6\u4e0e\u8f93\u5165\u5e8f\u5217\u76f8\u540c\uff0c\u5373ViT\u7684\u8f93\u51fa\u662f\u5355\u5c3a\u5ea6\u7684\uff1b<\/p>\n\n\n\n

(2)\u7531\u4e8e\u8d44\u6e90\u6709\u9650\uff0cViT\u7684\u8f93\u5165\u662f\u7c97\u7c92\u5ea6\u7684\uff08\u56fe\u50cf\u5757\u5927\u5c0f\u4e3a16\u621632\u50cf\u7d20\uff09\uff0c\u56e0\u6b64\u5176\u8f93\u51fa\u5206\u8fa8\u7387\u76f8\u5bf9\u8f83\u4f4e\uff1b<\/p>\n\n\n\n

(3)\u5373\u4f7f\u5bf9\u4e8e\u5e38\u89c1\u7684\u8f93\u5165\u56fe\u50cf\u5927\u5c0f\u5176\u8ba1\u7b97\u548c\u5185\u5b58\u6210\u672c\u76f8\u5bf9\u8f83\u9ad8\u3002<\/p>\n\n\n\n

\u4e3a\u4e86\u89e3\u51b3\u4e0a\u8ff0\u9650\u5236\uff0c\u672c\u5de5\u4f5c\u63d0\u51fa\u4e86\u4e00\u4e2a\u7eaf\u7cb9\u7684Transformer\u4e3b\u5e72\u7f51\u7edcPVT\uff0c\u901a\u8fc7\u5f15\u5165\u4e00\u4e2a\u6e10\u8fdb\u6536\u7f29\u91d1\u5b57\u5854\u548c\u7b80\u5355\u4f46\u6709\u6548\u7684\u6ce8\u610f\u5c42SRA\uff0c\u5904\u7406\u9ad8\u5206\u8fa8\u7387\u7684\u7279\u5f81\u56fe\uff0c\u751f\u6210\u591a\u5c3a\u5ea6\u7684\u7279\u5f81\u56fe\uff0c\u5e76\u964d\u4f4e\u8ba1\u7b97\/\u5185\u5b58\u6210\u672c\u3002\u5b83\u53ef\u4ee5\u5728\u8bb8\u591a\u4e0b\u6e38\u4efb\u52a1\u4e2d\u4f5c\u4e3aCNN\u4e3b\u5e72\u7684\u66ff\u4ee3\u65b9\u6848\uff0c\u5305\u62ec\u56fe\u50cf\u7ea7\u9884\u6d4b\u548c\u50cf\u7d20\u7ea7\u5bc6\u96c6\u9884\u6d4b\u3002<\/p>\n\n\n\n

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

\u4e8c\u3001Pyramid Vision Transformer<\/h2>\n\n\n\n

\u4e0eCNN\u9aa8\u5e72\u7c7b\u4f3c\uff0cPVT\u7684\u7279\u5f81\u91d1\u5b57\u5854\u6709\u56db\u4e2a\u9636\u6bb5\u6765\u751f\u6210\u4e0d\u540c\u5c3a\u5ea6\u7684\u7279\u5f81\u56fe\uff0c\u6bcf\u4e2a\u9636\u6bb5\u7531\u4e00\u4e2a\u56fe\u50cf\u5757\u5d4c\u5165\u5c42\u548c\u4e00\u4e2aL_i\u00a0 Transformer\u7f16\u7801\u5668\u5c42\u7ec4\u6210\uff0c\u4f7f\u7528\u4e00\u79cd\u6e10\u8fdb\u6536\u7f29\u7b56\u7565\uff0c\u901a\u8fc7\u56fe\u50cf\u5d4c\u5165\u5c42\u6765\u63a7\u5236\u7279\u5f81\u56fe\u7684\u5c3a\u5ea6\u3002<\/p>\n\n\n\n

\u2022\u9996\u5148\u5c06\u8f93\u5165\u7279\u5f81\u56feF_(i-1)\u2208R^(H_(i-1)\u00d7W_(i-1)\u00d7C_(i-1) )\u5e73\u5747\u5212\u5206\u4e3a(H_(i-1) W_(i-1))\/(P_i^2 )\u56fe\u50cf\u5757\uff08 P_i\u4e3a\u7b2ci\u4e2a\u9636\u6bb5\u7684\u56fe\u50cf\u5757\u5927\u5c0f\uff09\u3002<\/p>\n\n\n\n

\u2022\u7136\u540e\u628a\u6bcf\u4e2a\u56fe\u50cf\u5757\u53d8\u5e73\u5e76\u6295\u5f71\u5230C_i\u7ef4\u5d4c\u5165\u3002<\/p>\n\n\n\n

\u2022\u7ecf\u8fc7\u7ebf\u6027\u6295\u5f71\u540e\uff0c\u5d4c\u5165\u7684\u56fe\u50cf\u5757\u7684\u5f62\u72b6\u53d8\u4e3aH_(i-1)\/P_i \u00d7W_(i-1)\/P_i   \u00d7C_i\uff0c\u5176\u4e2d\u9ad8\u5ea6\u548c\u5bbd\u5ea6\u662f\u8f93\u5165\u7684P_i\u5206\u4e4b\u4e00\u3002<\/p>\n\n\n\n

\u4ee5\u6b64\u65b9\u5f0f\u53ef\u4ee5\u7075\u6d3b\u5730\u8c03\u6574\u6bcf\u4e2a\u9636\u6bb5\u7279\u5f81\u56fe\u7684\u6bd4\u4f8b\uff0c\u6784\u9020\u4e00\u4e2a\u7279\u5f81\u91d1\u5b57\u5854 {F_1 \u3001F_2 \u3001F_3 \u3001F_4}\uff0c\u5b83\u4eec\u76f8\u5bf9\u4e8e\u8f93\u5165\u56fe\u50cf\u7684\u6b65\u5e45\u5206\u522b\u4e3a4\u30018\u300116\u548c32\u50cf\u7d20\u3002 \u501f\u52a9\u7279\u5f81\u91d1\u5b57\u5854\uff0cPVT\u53ef\u4ee5\u5f88\u5bb9\u6613\u5730\u5e94\u7528\u4e8e\u5927\u591a\u6570\u4e0b\u6e38\u4efb\u52a1\uff0c\u5305\u62ec\u56fe\u50cf\u5206\u7c7b\u3001\u76ee\u6807\u68c0\u6d4b\u548c\u8bed\u4e49\u5206\u5272\u3002<\/p>\n\n\n\n

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

\u7531\u4e8ePVT\u9700\u8981\u5904\u7406\u9ad8\u5206\u8fa8\u7387\u7279\u5f81\u56fe\uff0c\u672c\u6587\u8fd8\u63d0\u51fa\u4e86\u4e00\u4e2a\u7a7a\u95f4\u51cf\u5c11\u6ce8\u610f\uff08SRA\uff09\u5c42\u6765\u53d6\u4ee3\u4f20\u7edf\u7f16\u7801\u5668\u4e2d\u7684\u591a\u5934\u6ce8\u610f\uff08MHA\uff09\u5c42\u3002\u4e0eMHA\u7c7b\u4f3c\uff0cSRA\u63a5\u6536\u4e00\u4e2a\u67e5\u8be2Q\u3001\u4e00\u4e2a\u5bc6\u94a5K\u548c\u4e00\u4e2a\u503cV\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u8f93\u51fa\u4e00\u4e2a\u6539\u8fdb\u7684\u7279\u5f81\u3002\u4e0d\u540c\u7684\u662f\uff0cSRA\u5728\u6ce8\u610f\u64cd\u4f5c\u524d\uff0c\u964d\u4f4e\u4e86K\u548cV\u7684\u7a7a\u95f4\u5c3a\u5ea6\uff0c\u8fd9\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u51cf\u5c11\u4e86\u8ba1\u7b97\u5f00\u9500\u3002<\/p>\n\n\n\n

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

\u4e09\u3001\u5b9e\u9a8c\u8bc4\u4f30<\/h2>\n\n\n\n
    \n
  • \u5728\u56fe\u50cf\u5206\u7c7b\u4e2d\uff0c\u672c\u6587\u8bbe\u8ba1\u4e86\u4e00\u7cfb\u5217\u4e0d\u540c\u5c3a\u5ea6\u7684PVT\u6a21\u578b\uff0c\u5373PVT-Tiny\u3001-Small, -Medium,\u548c-Large\uff0c\u5176\u53c2\u6570\u6570\u5206\u522b\u4e0eResNet18\u300150\u3001101\u548c152\u76f8\u4f3c\uff0c\u5e76\u9075\u5faaViT\u548cDeiT\uff0c\u5728\u6700\u540e\u4e00\u4e2a\u9636\u6bb5\u7684\u8f93\u5165\u4e2d\u6dfb\u52a0\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684\u5206\u7c7b\u4ee4\u724c\uff0c\u7136\u540e\u4f7f\u7528\u4e00\u4e2a\u5168\u8fde\u63a5\uff08FC\uff09\u5c42\u5728\u4ee4\u724c\u4e4b\u4e0a\u8fdb\u884c\u5206\u7c7b\u3002\u5b9e\u9a8c\u7ed3\u679c\u8bc1\u660e\uff1a\u5728\u76f8\u4f3c\u7684\u53c2\u6570\u6570\u548c\u8ba1\u7b97\u9884\u7b97\u4e0b\uff0cPVT\u6a21\u578b\u4f18\u4e8e\u4f20\u7edf\u7684CNN\u9aa8\u5e72\u3002\u540c\u65f6\uff0c\u5728\u7c7b\u4f3c\u6216\u66f4\u4f4e\u7684\u590d\u6742\u5ea6\u4e0b\uff0cPVT\u6a21\u578b\u7684\u6027\u80fd\u4e0e\u6700\u8fd1\u63d0\u51fa\u7684\u57fa\u4e8eTransformer\u7684\u6a21\u578b\u76f8\u5f53\u3002<\/li>\n<\/ul>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
      \n
    • \u5728\u76ee\u6807\u68c0\u6d4b\u4e2d\uff0c\u5bf9\u4e8eRetinaNet\uff0c\u5728\u76f8\u540c\u6570\u91cf\u7684\u53c2\u6570\u4e0b\uff0c\u57fa\u4e8ePVT\u7684\u6a21\u578b\u663e\u8457\u8d85\u8fc7\u4e86\u5176\u5bf9\u5e94\u6a21\u578b\uff0c\u4e14\u5728\u57fa\u4e8eMask R-CNN\u7684\u5b9e\u4f8b\u5206\u5272\u5b9e\u9a8c\u4e2d\u4e5f\u5f97\u5230\u4e86\u7c7b\u4f3c\u7684\u7ed3\u679c\u3002\u8fd9\u4e9b\u7ed3\u679c\u8868\u660e\uff0cPVT\u53ef\u4ee5\u4f5c\u4e3a\u4e00\u4e2a\u5f88\u597d\u7684\u66ff\u4ee3CNN\u9aa8\u5e72\u7684\u76ee\u6807\u68c0\u6d4b\u3002<\/li>\n<\/ul>\n\n\n\n
      \"\"<\/figure>\n\n\n\n
        \n
      • \u5728\u8bed\u4e49\u5206\u5272\u4e2d\uff0c\u5f53\u4f7f\u7528\u8bed\u4e49FPN\u65f6\uff0c\u57fa\u4e8ePVT\u7684\u6a21\u578b\u7684\u6027\u80fd\u59cb\u7ec8\u4f18\u4e8e\u57fa\u4e8eResNet\u6216ResNeXt\u7684\u6a21\u578b\u3002<\/li>\n<\/ul>\n\n\n\n
        \"\"<\/figure>\n\n\n\n

        \u56db\u3001\u603b\u7ed3<\/h2>\n\n\n\n

        \u672c\u6587\u63d0\u51fa\u7684\u91d1\u5b57\u5854\u89c6\u89c9Transformer\uff08PVT\uff09\u662f\u7b2c\u4e00\u4e2a\u4e3a\u5404\u79cd\u50cf\u7d20\u7ea7\u5bc6\u96c6\u9884\u6d4b\u4efb\u52a1\u8bbe\u8ba1\u7684\u7eafTransformer\u4e3b\u5e72\u3002\u901a\u8fc7\u8bbe\u8ba1\u4e00\u4e2a\u6e10\u8fdb\u5f0f\u6536\u7f29\u91d1\u5b57\u5854\u548c\u4e00\u4e2a\u7a7a\u95f4\u7f29\u51cf\u6ce8\u610f\u5c42\uff08SRA\uff09\uff0c\u83b7\u5f97\u4e86\u9ad8\u5206\u8fa8\u7387\u548c\u591a\u5c3a\u5ea6\u7684\u7279\u5f81\u56fe\uff0c\u51cf\u5c11\u4e86Transformer\u7684\u8d44\u6e90\u6d88\u8017\uff0c\u514b\u670d\u4e86\u5c06Transformer\u79fb\u690d\u5230\u5bc6\u96c6\u9884\u6d4b\u65f6\u7684\u8bb8\u591a\u56f0\u96be\u3002\u5728\u5bf9\u8c61\u68c0\u6d4b\u548c\u8bed\u4e49\u5206\u5272\u57fa\u51c6\u4e0a\u7684\u5927\u91cf\u5b9e\u9a8c\u8bc1\u660e\uff0c\u5728\u76f8\u540c\u6570\u91cf\u7684\u53c2\u6570\u4e0b\uff0cPVT\u6bd4\u8bbe\u8ba1\u826f\u597d\u7684CNN\u9aa8\u5e72\u66f4\u5f3a\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

        \u6765\u6e90\uff1a2021 IEEE\/CVF International Conference on Computer V […]<\/p>\n","protected":false},"author":19,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-789","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/789","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/users\/19"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=789"}],"version-history":[{"count":4,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/789\/revisions"}],"predecessor-version":[{"id":928,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/789\/revisions\/928"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=789"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=789"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=789"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}