{"id":12350,"date":"2023-02-04T10:54:35","date_gmt":"2023-02-04T02:54:35","guid":{"rendered":"http:\/\/139.9.1.231\/?p=12350"},"modified":"2023-02-04T10:54:37","modified_gmt":"2023-02-04T02:54:37","slug":"mpvit-multi-path-vision-transformer-for-dense-prediction","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2023\/02\/04\/mpvit-multi-path-vision-transformer-for-dense-prediction\/","title":{"rendered":"MPViT : Multi-Path Vision Transformer for Dense Prediction"},"content":{"rendered":"\n<figure class=\"wp-block-pullquote\"><blockquote><p>https:\/\/arxiv.org\/abs\/2112.11010\uff08CVPR2022\uff09<\/p><cite>https:\/\/github.com\/youngwanLEE\/MPViT<\/cite><\/blockquote><\/figure>\n\n\n\n<h3 id=\"h_478325208_0\">Introduction<\/h3>\n\n\n\n<ul><li>\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\uff0c\u4f5c\u8005\u4ee5\u4e0d\u540c\u4e8e\u73b0\u6709Transformer\u7684\u89c6\u89d2\uff0c\u63a2\u7d22\u591a\u5c3a\u5ea6path embedding\u4e0emulti-path\u7ed3\u6784\uff0c\u63d0\u51fa\u4e86<strong>Multi-path Vision Transformer(MPViT)<\/strong>\u3002<\/li><li>\u901a\u8fc7\u4f7f\u7528 overlapping convolutional\u00a0<strong>patch embedding<\/strong>\uff0cMPViT\u540c\u65f6\u5d4c\u5165\u76f8\u540c\u5927\u5c0f\u7684patch\u7279\u5f81\u3002\u7136\u540e\uff0c\u5c06\u4e0d\u540c\u5c3a\u5ea6\u7684Token\u901a\u8fc7\u591a\u6761\u8def\u5f84\u72ec\u7acb\u5730\u8f93\u5165<strong>Transformer encoders<\/strong>\uff0c\u5e76\u5bf9\u751f\u6210\u7684\u7279\u5f81\u8fdb\u884c\u805a\u5408\uff0c\u4ece\u800c\u5728\u540c\u4e00\u7279\u5f81\u7ea7\u522b\u4e0a\u5b9e\u73b0\u7cbe\u7ec6\u548c\u7c97\u7cd9\u7684\u7279\u5f81\u8868\u793a\u3002<\/li><li>\u5728\u7279\u5f81\u805a\u5408\u6b65\u9aa4\u4e2d\uff0c\u5f15\u5165\u4e86\u4e00\u4e2a<strong>global-to-local feature interaction(GLI)<\/strong>\u8fc7\u7a0b\uff0c\u8be5\u8fc7\u7a0b\u5c06\u5377\u79ef\u5c40\u90e8\u7279\u5f81\u4e0eTransformer\u7684\u5168\u5c40\u7279\u5f81\u8fde\u63a5\u8d77\u6765\uff0c\u540c\u65f6\u5229\u7528\u4e86\u5377\u79ef\u7684\u5c40\u90e8\u8fde\u901a\u6027\u548cTransformer\u7684\u5168\u5c40\u4e0a\u4e0b\u6587\u3002<\/li><\/ul>\n\n\n\n<h3><strong>Vision Transformers for dense predictions<\/strong>.<\/h3>\n\n\n\n<ul><li>\u5bc6\u96c6\u7684\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\uff0c\u5982\u76ee\u6807\u68c0\u6d4b\u548c\u5206\u5272\uff0c\u9700\u8981\u6709\u6548\u7684\u591a\u5c3a\u5ea6\u7279\u5f81\u8868\u793a\uff0c\u4ee5\u68c0\u6d4b\u6216\u5206\u7c7b\u4e0d\u540c\u5927\u5c0f\u7684\u7269\u4f53\u6216\u533a\u57df\u3002Vision Transformer(ViT)\u6784\u5efa\u4e86\u4e00\u4e2a\u7b80\u5355\u7684\u591a\u9636\u6bb5\u7ed3\u6784\uff08\u5373\u7cbe\u7ec6\u5230\u7c97\u7cd9\uff09\uff0c\u7528\u4e8e\u4f7f\u7528\u5355\u5c3a\u5ea6patch\u7684\u591a\u5c3a\u5ea6\u8868\u793a\u3002\u7136\u800cViT\u7684\u53d8\u4f53\u4e13\u6ce8\u4e8e\u964d\u4f4e\u81ea\u6ce8\u610f\u7684\u4e8c\u6b21\u590d\u6742\u5ea6\uff0c\u8f83\u5c11\u5173\u6ce8\u6784\u5efa\u6709\u6548\u7684\u591a\u5c3a\u5ea6\u8868\u793a\u3002<\/li><li>CoaT\u901a\u8fc7\u4f7f\u7528\u4e00\u79cdco-scale\u673a\u5236\uff0c\u540c\u65f6\u8868\u793a\u7cbe\u7ec6\u548c\u7c97\u7cd9\u7684\u7279\u5f81\uff0c\u5141\u8bb8\u5e76\u884c\u5730\u8de8\u5c42\u6ce8\u610f\uff0c\u4ece\u800c\u63d0\u9ad8\u4e86\u68c0\u6d4b\u6027\u80fd\u3002\u7136\u800c\uff0cco-scale\u673a\u5236\u9700\u8981\u5927\u91cf\u7684\u8ba1\u7b97\u548c\u5185\u5b58\u5f00\u9500\uff0c\u56e0\u4e3a\u5b83\u4e3a\u57fa\u7840\u6a21\u578b\u589e\u52a0\u4e86\u989d\u5916\u7684\u8de8\u5c42\u5173\u6ce8(\u4f8b\u5982\uff0cCoaT-Lite)\u3002\u56e0\u6b64\uff0c\u5bf9\u4e8eViT\u4f53\u7cfb\u7ed3\u6784\u7684\u591a\u5c3a\u5ea6\u7279\u5f81\u8868\u793a\u4ecd\u6709\u6539\u8fdb\u7684\u7a7a\u95f4\u3002<\/li><\/ul>\n\n\n\n<h3><strong>Comparison to Concurrent work.<\/strong><\/h3>\n\n\n\n<ul><li>CrossViT\u5229\u7528\u4e86\u4e0d\u540c\u7684patch\u5927\u5c0f\u548c\u5355\u7ea7\u7ed3\u6784\u4e2d\u7684\u53cc\u8def\u5f84\uff0c\u5982ViT\u548cXCiT\u3002\u7136\u800c\uff0cCrossViT\u7684\u5206\u652f\u4e4b\u95f4\u7684\u76f8\u4e92\u4f5c\u7528\u53ea\u901a\u8fc7[CLS]token\u53d1\u751f\uff0c\u800cMPViT\u5141\u8bb8\u6240\u6709\u4e0d\u540c\u89c4\u6a21\u7684patch\u76f8\u4e92\u4f5c\u7528\u3002\u6b64\u5916\uff0c\u4e0eCrossViT\uff08\u4ec5\u9650\u5206\u7c7b\uff09\u4e0d\u540c\u7684\u662f\uff0cMPViT\u66f4\u666e\u904d\u5730\u63a2\u7d22\u66f4\u5927\u7684\u8def\u5f84\u7ef4\u5ea6\uff08\u4f8b\u5982\uff0c\u8d85\u8fc7\u4e24\u4e2a\u7ef4\u5ea6\uff09\uff0c\u5e76\u91c7\u7528\u591a\u9636\u6bb5\u7ed3\u6784\u8fdb\u884c\u5bc6\u96c6\u9884\u6d4b\u3002<\/li><\/ul>\n\n\n\n<h3>Method<\/h3>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"565\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/01\/image-94-1024x565.png\" alt=\"\" class=\"wp-image-12352\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/01\/image-94-1024x565.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/01\/image-94-300x165.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/01\/image-94-768x423.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/01\/image-94.png 1453w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"515\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/01\/image-95-1024x515.png\" alt=\"\" class=\"wp-image-12354\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/01\/image-95-1024x515.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/01\/image-95-300x151.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/01\/image-95-768x386.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/01\/image-95.png 1380w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h4 id=\"h_478325208_2\"><strong>Conv-stem<\/strong><\/h4>\n\n\n\n<p>\u8f93\u5165\u56fe\u50cf\u5927\u5c0f\u4e3a\uff1aH\u00d7W\u00d73\uff0c\u4e24\u5c42\u5377\u79ef\uff1a\u91c7\u7528\u4e24\u4e2a3\u00d73\u7684\u5377\u79ef\uff0c\u901a\u9053\u5206\u522b\u4e3aC2\/2\uff0cC2\uff0cstride\u4e3a2\uff0c\u751f\u6210\u7279\u5f81\u7684\u5927\u5c0f\u4e3aH\/4\u00d7W\/4\u00d7C2\uff0c\u5176\u4e2dC2\u4e3astage 2\u7684\u901a\u9053\u5927\u5c0f\u3002<\/p>\n\n\n\n<ul><li>\u8bf4\u660e\uff1a\u6bcf\u4e2a\u5377\u79ef\u4e4b\u540e\u90fd\u662fBatch Normalization \u548c\u4e00\u4e2aHardswish\u6fc0\u6d3b\u51fd\u6570\u3002<\/li><li>In LeViT , a convolutional stem block shows better low-level representation (<em>i.e<\/em>., without losing salient information) than non-overlapping patch embedding.<\/li><\/ul>\n\n\n\n<p>\u4ecestage 2\u5230stage 5\uff0c\u4f5c\u8005\u5728\u6bcf\u4e2a\u9636\u6bb5\u5bf9\u6240\u63d0\u51fa\u7684<strong>Multi-scale Patch Embedding<\/strong>(MS-PatchEmbed)\u548c<strong>Multi-path Transformer<\/strong>(MP-Transformer)\u5757\u8fdb\u884c\u5806\u53e0<strong>\u3002<\/strong><\/p>\n\n\n\n<h4><strong>Multi-Scale Patch Embedding<\/strong><\/h4>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"980\" height=\"601\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-31.png\" alt=\"\" class=\"wp-image-12771\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-31.png 980w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-31-300x184.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-31-768x471.png 768w\" sizes=\"(max-width: 980px) 100vw, 980px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"214\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-32-1024x214.png\" alt=\"\" class=\"wp-image-12772\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-32-1024x214.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-32-300x63.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-32-768x161.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-32.png 1090w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u901a\u8fc7\u6539\u53d8stride\u548cpadding\u6765\u8c03\u6574token\u7684\u5e8f\u5217\u957f\u5ea6\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u53ef\u4ee5\u8f93\u51fa\u5177\u6709\u4e0d\u540cpatch\u5927\u5c0f\u7684\u76f8\u540c\u5927\u5c0f\uff08\u5373\u5206\u8fa8\u7387\uff09\u7684\u7279\u5f81\u3002\u56e0\u6b64\uff0c\u4f5c\u8005\u5e76\u884c\u5730\u5f62\u6210\u4e86\u51e0\u4e2a\u5177\u6709\u4e0d\u540c\u5377\u79ef\u6838\u5927\u5c0f\u7684\u5377\u79efpatch embedding\u5c42\u3002\u4f8b\u5982\uff0c\u5982\u56fe1\u6240\u793a\uff0c\u53ef\u4ee5\u751f\u6210\u76f8\u540c\u5e8f\u5217\u957f\u5ea6\u7684\u4e0d\u540c\u5927\u5c0f\u7684vision token\uff0cpatch\u5927\u5c0f\u5206\u522b\u4e3a3\u00d73,5\u00d75,7\u00d77\u3002<\/p>\n\n\n\n<p>\u7531\u4e8e\u5177\u6709\u76f8\u540c\u901a\u9053\u548c\u6ee4\u6ce2\u5668\u5927\u5c0f\u7684\u8fde\u7eed\u5377\u79ef\u64cd\u4f5c\u6269\u5927\u4e86\u63a5\u53d7\u57df\uff0c\u5e76\u4e14\u9700\u8981\u66f4\u5c11\u7684\u53c2\u6570\uff0c\u5728\u5b9e\u8df5\u4e2d\u9009\u62e9\u4e86\u8fde\u7eed\u76843\u00d73\u5377\u79ef\u5c42\u3002\u4e3a\u4e86\u51cf\u5c11\u53c2\u6570\u91cf\uff0c\u5728\u5b9e\u8df5\u4e2d\u9009\u62e9\u4e86\u4e24\u4e2a\u8fde\u7eed\u76843\u00d73\u5377\u79ef\u5c42\u4ee3\u66ff5\u00d75\u5377\u79ef\u3002\u5bf9\u4e8etriple-path\u7ed3\u6784\uff0c\u4f7f\u7528\u4e09\u4e2a\u8fde\u7eed\u76843\u00d73\u5377\u79ef\uff0c\u901a\u9053\u5927\u5c0f\u4e3aC&#8217;\uff0cpadding\u4e3a1\uff0c\u6b65\u5e45\u4e3as\uff0c\u5176\u4e2ds\u5728\u964d\u4f4e\u7a7a\u95f4\u5206\u8fa8\u7387\u65f6\u4e3a2\uff0c\u5426\u5219\u4e3a1\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"127\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-33-1024x127.png\" alt=\"\" class=\"wp-image-12775\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-33-1024x127.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-33-300x37.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-33-768x95.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-33.png 1065w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<ul><li>\u8bf4\u660e\uff1a\u4e3a\u4e86\u51cf\u5c11\u6a21\u578b\u53c2\u6570\u548c\u8ba1\u7b97\u5f00\u9500\uff0c\u91c7\u75283\u00d73\u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef\uff0c\u5305\u62ec3\u00d73\u6df1\u5ea6\u5377\u79ef\u548c1\u00d71\u70b9\u5377\u79ef\u3002<\/li><li>\u6bcf\u4e2a\u5377\u79ef\u4e4b\u540e\u90fd\u662fBatch Normalization \u548c\u4e00\u4e2aHardswish\u6fc0\u6d3b\u51fd\u6570\u3002<\/li><\/ul>\n\n\n\n<p>\u63a5\u7740\uff0c\u4e0d\u540c\u5927\u5c0f\u7684token embedding features \u5206\u522b\u8f93\u5165\u5230transformer encoder\u4e2d\u3002<\/p>\n\n\n\n<h4>Multi-path Transformer<\/h4>\n\n\n\n<p><strong>\u539f\u56e0\uff1a<\/strong>Transformer\u4e2d\u7684self-attention\u53ef\u4ee5\u6355\u83b7\u957f\u671f\u4f9d\u8d56\u5173\u7cfb\uff08\u5373\u5168\u5c40\u4e0a\u4e0b\u6587\uff09\uff0c\u4f46\u5b83\u5f88\u53ef\u80fd\u4f1a\u5ffd\u7565\u6bcf\u4e2apatch\u4e2d\u7684\u7ed3\u6784\u6027\u4fe1\u606f\u548c\u5c40\u90e8\u5173\u7cfb\u3002\u76f8\u53cd\uff0ccnn\u53ef\u4ee5\u5229\u7528\u5e73\u79fb\u4e0d\u53d8\u6027\u4e2d\u7684\u5c40\u90e8\u8fde\u901a\u6027\uff0c\u4f7f\u5f97CNN\u5728\u5bf9\u89c6\u89c9\u5bf9\u8c61\u8fdb\u884c\u5206\u7c7b\u65f6\uff0c\u5bf9\u7eb9\u7406\u6709\u66f4\u5f3a\u7684\u4f9d\u8d56\u6027\uff0c\u800c\u4e0d\u662f\u5f62\u72b6\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0cMPViT\u4ee5\u4e00\u79cd\u4e92\u8865\u7684\u65b9\u5f0f\u5c06CNN\u4e0eTransformer\u7ed3\u5408\u8d77\u6765\u3002<\/p>\n\n\n\n<ul><li>\u4e3a\u4e86\u8868\u793a\u5c40\u90e8\u7279\u5f81\u00a0<\/li><\/ul>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-34.png\" alt=\"\" class=\"wp-image-12779\" width=\"123\" height=\"20\"\/><\/figure><\/div>\n\n\n\n<ul><li>\u91c7\u7528\u4e86\u4e00\u4e2a depthwise residual bottleneck block\uff0c\u5305\u62ec1\u00d71\u5377\u79ef\u30013\u00d73\u6df1\u5ea6\u5377\u79ef\u548c1\u00d71\u5377\u79ef\u548c\u6b8b\u5dee\u8fde\u63a5\u3002<\/li><li>\u4e3a\u4e86\u51cf\u8f7b\u591a\u8def\u5f84\u7ed3\u6784\u7684\u8ba1\u7b97\u8d1f\u62c5\uff0c\u4f7f\u7528\u4e86CoaT\u4e2d\u63d0\u51fa\u7684\u6709\u6548\u7684\u56e0\u7d20\u5206\u89e3\u81ea\u6ce8\u610f\uff1a<\/li><\/ul>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"631\" height=\"105\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-35.png\" alt=\"\" class=\"wp-image-12782\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-35.png 631w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-35-300x50.png 300w\" sizes=\"(max-width: 631px) 100vw, 631px\" \/><\/figure>\n\n\n\n<h4><strong>Global-to-Local Feature Interaction<\/strong><\/h4>\n\n\n\n<p>\u5c06\u5c40\u90e8\u7279\u5f81\u548c\u5168\u5c40\u7279\u5f81\u805a\u5408\u8d77\u6765\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"214\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-36-1024x214.png\" alt=\"\" class=\"wp-image-12784\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-36-1024x214.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-36-300x63.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-36-768x160.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-36.png 1083w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u4e3a\u4e86\u4fdd\u6301\u53ef\u6bd4\u6027\u7684\u53c2\u6570\u548cFLOPs\uff0c\u589e\u52a0\u8def\u5f84\u7684\u6570\u91cf\u9700\u8981\u51cf\u5c11\u901a\u9053C\u6216\u5c42\u6570L\uff08\u5373\uff0ctransformer encoder\u7684\u6570\u91cf\uff09\u3002\u4f5c\u8005\u901a\u8fc7\u51cf\u5c11C\u800c\u4e0d\u662fL\uff0c\u4ece\u5355\u8def\u5f84(\u5373CoaT-Lite baseline)\u6269\u5c55\u5230triple-path\u3002\u5728\u6d88\u878d\u7814\u7a76\u4e2d\uff0c\u9a8c\u8bc1\u4e86\u51cf\u5c11C\u6bd4\u51cf\u5c11L\u83b7\u5f97\u66f4\u597d\u7684\u6027\u80fd\uff08\u89c1\u88685\uff09\u3002\u7531\u4e8estage2\u7684\u7279\u5f81\u5206\u8fa8\u7387\u8f83\u9ad8\uff0c\u5bfc\u81f4\u8ba1\u7b97\u6210\u672c\u8f83\u9ad8\uff0c\u4f5c\u8005\u5728stage2\u4e2d\u5c06triple-path\u6a21\u578b\u7684\u8def\u5f84\u6570\u8bbe\u7f6e\u4e3a2\u3002\u4ecestage3\u5f00\u59cb\uff0c\u4e09\u8def\u5f84\u6a21\u578b\u67093\u6761\u8def\u5f84\u3002<\/p>\n\n\n\n<p>\u4f5c\u8005\u8fd8\u53d1\u73b0\uff0c\u867d\u7136 triple-path\u548c\u53cc\u8def\u5f84\u5728ImageNet\u5206\u7c7b\u4e2d\u4ea7\u751f\u76f8\u4f3c\u7684\u7cbe\u5ea6\uff0c\u4f46 triple-path\u6a21\u578b\u5728\u5bc6\u96c6\u9884\u6d4b\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u66f4\u597d\u7684\u6027\u80fd\u3002\u56e0\u6b64\uff0c\u5efa\u7acb\u4e86\u57fa\u4e8e triple-path\u7ed3\u6784\u7684MPViT\u6a21\u578b\u3002MPViT\u7684\u8be6\u7ec6\u60c5\u51b5\u89c1\u88681\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"712\" height=\"156\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-37.png\" alt=\"\" class=\"wp-image-12785\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-37.png 712w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-37-300x66.png 300w\" sizes=\"(max-width: 712px) 100vw, 712px\" \/><\/figure>\n\n\n\n<h3 id=\"h_478325208_5\">Experiments<\/h3>\n\n\n\n<h4 id=\"h_478325208_6\"><strong>Ablation study<\/strong><\/h4>\n\n\n\n<p>\u5bf9MPViT-XS\u7684\u6bcf\u4e2a\u7ec4\u4ef6\u8fdb\u884c\u6d88\u878d\u7814\u7a76\uff0c\u4ee5\u7814\u7a76\u63d0\u51fa\u7684\u591a\u8def\u5f84\u7ed3\u6784\u5bf9\u56fe\u50cf\u5206\u7c7b\u548c\u4f7f\u7528Mask R-CNN\u68c0\u6d4b\u7684\u6709\u6548\u6027\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"215\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-38-1024x215.png\" alt=\"\" class=\"wp-image-12787\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-38-1024x215.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-38-300x63.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-38-768x161.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-38.png 1461w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"750\" height=\"256\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-39.png\" alt=\"\" class=\"wp-image-12788\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-39.png 750w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-39-300x102.png 300w\" sizes=\"(max-width: 750px) 100vw, 750px\" \/><\/figure>\n\n\n\n<p><strong>Exploring path dimension<\/strong>.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"714\" height=\"386\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-40.png\" alt=\"\" class=\"wp-image-12790\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-40.png 714w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-40-300x162.png 300w\" sizes=\"(max-width: 714px) 100vw, 714px\" \/><\/figure>\n","protected":false},"excerpt":{"rendered":"<p>https:\/\/arxiv.org\/abs\/2112.11010\uff08CVPR2022\uff09 https:\/\/gith &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2023\/02\/04\/mpvit-multi-path-vision-transformer-for-dense-prediction\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">MPViT : Multi-Path Vision Transformer for Dense Prediction<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[21,24,36,25,9],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/12350"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=12350"}],"version-history":[{"count":25,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/12350\/revisions"}],"predecessor-version":[{"id":12791,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/12350\/revisions\/12791"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=12350"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=12350"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=12350"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}