{"id":94061,"date":"2026-08-13T15:49:50","date_gmt":"2026-08-13T07:49:50","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/94061.html"},"modified":"2026-08-13T15:49:50","modified_gmt":"2026-08-13T07:49:50","slug":"%ef%bc%88%e8%ae%ba%e6%96%87%e9%80%9f%e8%af%bb%ef%bc%89ego%ef%bc%9a%e5%86%85%e5%ae%b9%e7%9b%b8%e4%bc%bc%e5%ba%a6%e9%a9%b1%e5%8a%a8%e7%9a%84%e5%8d%b7%e7%a7%af%e7%a5%9e%e7%bb%8f%e7%bd%91%e7%bb%9c","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/94061.html","title":{"rendered":"\uff08\u8bba\u6587\u901f\u8bfb\uff09Ego\uff1a\u5185\u5bb9\u76f8\u4f3c\u5ea6\u9a71\u52a8\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc"},"content":{"rendered":"<p style=\"margin-left:0;margin-right:0;text-align:left\">\u8bba\u6587\u9898\u76ee&#xff1a;Convolutional Neural Networks Driven by Content Similarity&#xff08;\u5185\u5bb9\u76f8\u4f3c\u5ea6\u9a71\u52a8\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc&#xff09;<\/p>\n<p style=\"margin-left:0;margin-right:0;text-align:left\">\u4f1a\u8bae&#xff1a;CVPR 2026<\/p>\n<p style=\"margin-left:0px;margin-right:0px;text-align:justify\">\u6458\u8981&#xff1a;\u867d\u7136\u5377\u79ef\u795e\u7ecf\u7f51\u7edc(CNN)\u8fd1\u5e74\u6765\u4e0d\u65ad\u53d1\u5c55&#xff0c;\u4f46\u53d8\u5f62\u91d1\u521a\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u53d8\u5f97\u8d8a\u6765\u8d8a\u53d7\u6b22\u8fce\u3002\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d&#xff0c;\u6211\u4eec\u4ecb\u7ecd\u4e86\u4e00\u79cd\u57fa\u4e8e\u5185\u5bb9\u76f8\u4f3c\u5ea6\u7684CNN\u4fe1\u606f\u805a\u5408\u673a\u5236-\u4e00\u79cd\u7c7b\u4f3c\u4e8e\u81ea\u6211\u6ce8\u610f\u673a\u5236\u7684\u80fd\u529b\u3002\u4e3a\u6b64&#xff0c;\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u6392\u5e8f\u64cd\u4f5c&#xff0c;\u5c06\u6807\u8bb0\u4e4b\u95f4\u7684\u7279\u5f81\u76f8\u4f3c\u5ea6\u8f6c\u6362\u4e3a\u76f8\u5bf9\u4f4d\u7f6e\u4fe1\u606f&#xff1a;\u5177\u4f53\u5730\u8bf4&#xff0c;\u91cd\u65b0\u6392\u5217\u6807\u8bb0&#xff0c;\u4f7f\u7279\u5f81\u76f8\u4f3c\u5ea6\u8f83\u9ad8\u7684\u6807\u8bb0\u66f4\u7d27\u5bc6\u5730\u653e\u7f6e\u5728\u4e00\u8d77\u3002\u8be5\u65b9\u6cd5\u5141\u8bb8\u5377\u79ef\u8fd0\u7b97\u95f4\u63a5\u8f6c\u6362\u4e3a\u7531\u5185\u5bb9\u76f8\u4f3c\u5ea6\u9a71\u52a8\u7684\u805a\u96c6\u6a21\u5f0f\u3002\u5b9e\u9a8c\u8868\u660e&#xff0c;\u6211\u4eec\u63d0\u51fa\u7684\u6a21\u578bEGO\u5728\u5404\u79cd\u4efb\u52a1\u4e2d\u53d6\u5f97\u4e86\u4ee4\u4eba\u6ee1\u610f\u7684\u7ed3\u679c&#xff0c;\u7a81\u663e\u4e86CNN\u5c1a\u672a\u5f00\u53d1\u7684\u6f5c\u529b\u3002<\/p>\n<p style=\"margin-left:0px;margin-right:0px;text-align:justify\">\u4ee3\u7801\u548c\u6a21\u578b\u53ef\u5728\u4ee5\u4e0b\u7f51\u5740\u516c\u5f00\u83b7\u53d6&#xff1a;https&#xff1a;\/\/githeb.com\/essenceoftheworld\/ego\u3002<\/p>\n<hr \/>\n<h3>\u4e00\u3001\u7814\u7a76\u80cc\u666f&#xff1a;CNN \u548c Self-Attention \u7684\u6838\u5fc3\u5dee\u5f02\u5230\u5e95\u5728\u54ea\u91cc&#xff1f;<\/h3>\n<p>\u5728 Vision Transformer \u51fa\u73b0\u4e4b\u524d&#xff0c;\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u957f\u671f\u5360\u636e\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u4e3b\u6d41\u4f4d\u7f6e\u3002\u4ece AlexNet\u3001VGG\u3001ResNet&#xff0c;\u5230\u540e\u6765\u7684 MobileNet\u3001ShuffleNet&#xff0c;CNN \u51ed\u501f\u5c40\u90e8\u8fde\u63a5\u3001\u53c2\u6570\u5171\u4eab\u548c\u826f\u597d\u7684\u8ba1\u7b97\u6548\u7387\u5efa\u7acb\u4e86\u975e\u5e38\u6210\u719f\u7684\u89c6\u89c9\u5efa\u6a21\u4f53\u7cfb\u3002\u968f\u540e SENet\u3001CBAM \u7b49\u5de5\u4f5c\u53c8\u5c06 Attention \u5f15\u5165 CNN&#xff0c;\u5bf9 channel \u6216 spatial feature \u8fdb\u884c\u52a8\u6001\u52a0\u6743&#xff0c;\u4f7f CNN \u7684\u8868\u8fbe\u80fd\u529b\u8fdb\u4e00\u6b65\u589e\u5f3a\u3002<\/p>\n<p>\u4f46\u662f ViT \u51fa\u73b0\u4ee5\u540e&#xff0c;\u89c6\u89c9\u6a21\u578b\u7684\u4e00\u4e2a\u91cd\u8981\u53d1\u5c55\u65b9\u5411\u53d1\u751f\u4e86\u53d8\u5316\u3002Self-Attention \u4e0d\u518d\u53ea\u4f9d\u8d56\u5c40\u90e8\u7a7a\u95f4\u4f4d\u7f6e\u8fdb\u884c\u4fe1\u606f\u805a\u5408&#xff0c;\u800c\u662f\u5141\u8bb8\u4efb\u610f\u4e24\u4e2a token \u6839\u636e\u5b83\u4eec\u7684\u5185\u5bb9\u5173\u7cfb\u76f4\u63a5\u53d1\u751f\u4ea4\u4e92&#xff0c;\u56e0\u6b64\u5728 long-range dependency \u548c content-dependent interaction \u4e0a\u8868\u73b0\u51fa\u4e86\u5f88\u5f3a\u7684\u7075\u6d3b\u6027\u3002<\/p>\n<p>\u4e3a\u4e86\u91cd\u65b0\u63d0\u9ad8 CNN \u7684\u7ade\u4e89\u529b&#xff0c;\u8fd1\u51e0\u5e74\u51fa\u73b0\u4e86\u5f88\u591a Transformer-style ConvNet&#xff0c;\u4f8b\u5982 ConvNeXt\u3001VAN\u3001Conv2Former\u3001HorNet\u3001MogaNet\u3001OverLoCK \u7b49\u3002\u8fd9\u4e9b\u65b9\u6cd5\u901a\u8fc7 large kernel\u3001gating mechanism\u3001high-order interaction \u7b49\u65b9\u5f0f\u4e0d\u65ad\u6269\u5927 CNN \u7684\u6709\u6548\u611f\u53d7\u91ce\u6216\u8005\u63d0\u9ad8 feature interaction \u80fd\u529b\u3002<\/p>\n<p>\u4f46\u662f\u4f5c\u8005\u8ba4\u4e3a&#xff0c;\u8fd9\u4e9b\u5de5\u4f5c\u867d\u7136\u4e0d\u65ad\u501f\u9274 Transformer \u7684\u8bbe\u8ba1\u601d\u60f3&#xff0c;\u4f20\u7edf convolution operation \u672c\u8eab\u7684\u6838\u5fc3\u903b\u8f91\u4ecd\u7136\u6ca1\u6709\u53d1\u751f\u53d8\u5316\u3002<\/p>\n<p>\u666e\u901a convolution \u7684\u4fe1\u606f\u805a\u5408\u4e3b\u8981\u7531relative spatial position\u51b3\u5b9a\u3002\u800c Self-Attention \u7684\u4fe1\u606f\u805a\u5408\u4e3b\u8981\u7531feature\/content similarity\u51b3\u5b9a\u3002<\/p>\n<p>\u8fd9\u5c31\u662f\u4f5c\u8005\u8ba4\u4e3a CNN \u548c Self-Attention \u4e4b\u95f4\u4e00\u4e2a\u975e\u5e38\u672c\u8d28\u7684\u673a\u5236\u5dee\u5f02\u3002<\/p>\n<p>\u5047\u8bbe\u4e2d\u5fc3\u4f4d\u7f6e\u5b58\u5728\u4e00\u4e2a token&#xff0c;\u800c\u53e6\u5916\u6709\u4e24\u4e2a token&#xff1a;<\/p>\n<p>Token A&#xff1a;\u7a7a\u95f4\u8ddd\u79bb\u975e\u5e38\u8fd1&#xff0c;\u4f46\u662f\u5185\u5bb9\u4e0e\u4e2d\u5fc3 token \u5dee\u5f02\u5f88\u5927&#xff1b;<\/p>\n<p>Token B&#xff1a;\u7a7a\u95f4\u8ddd\u79bb\u975e\u5e38\u8fdc&#xff0c;\u4f46\u662f\u8bed\u4e49\u6216\u8005 feature \u4e0e\u4e2d\u5fc3 token \u975e\u5e38\u76f8\u4f3c\u3002<\/p>\n<p>\u666e\u901a convolution \u66f4\u5bb9\u6613\u76f4\u63a5\u805a\u5408 Token A&#xff0c;\u56e0\u4e3a\u5b83\u4f4d\u4e8e\u5377\u79ef\u7a97\u53e3\u4e2d&#xff1b;Token B \u5373\u4f7f\u5185\u5bb9\u9ad8\u5ea6\u76f8\u5173&#xff0c;\u53ea\u8981\u8d85\u51fa\u5f53\u524d receptive field&#xff0c;\u5c31\u65e0\u6cd5\u76f4\u63a5\u53d1\u751f\u4fe1\u606f\u4ea4\u6362\u3002<\/p>\n<p>Self-Attention \u5219\u53ef\u4ee5\u901a\u8fc7 Query \u548c Key \u7684\u76f8\u4f3c\u7a0b\u5ea6\u76f4\u63a5\u5efa\u7acb\u4e2d\u5fc3 token \u4e0e Token B \u7684\u5173\u7cfb\u3002<\/p>\n<p>\u56e0\u6b64\u8bba\u6587\u63d0\u51fa\u7684\u95ee\u9898\u975e\u5e38\u76f4\u63a5&#xff1a;<\/p>\n<p>\u80fd\u4e0d\u80fd\u8ba9 convolution \u672c\u8eab\u4e5f\u62e5\u6709\u6839\u636e content similarity \u805a\u5408\u4fe1\u606f\u7684\u80fd\u529b&#xff0c;\u800c\u4e0d\u662f\u5fc5\u987b\u4f9d\u8d56\u663e\u5f0f Self-Attention&#xff1f;<\/p>\n<p>\u4f5c\u8005\u6700\u7ec8\u63d0\u51fa\u4e86\u4e00\u4e2a\u975e\u5e38\u6709\u610f\u601d\u7684\u7b54\u6848&#xff1a;<\/p>\n<p>\u4e0d\u7528\u76f4\u63a5\u6539\u53d8\u5377\u79ef&#xff0c;\u800c\u662f\u5148\u6539\u53d8 token \u7684\u6392\u5217\u987a\u5e8f\u3002<\/p>\n<hr \/>\n<h3>\u4e8c\u3001\u6838\u5fc3\u601d\u60f3&#xff1a;\u628a\u201c\u5185\u5bb9\u76f8\u4f3c\u201d\u8f6c\u6362\u6210\u201c\u4f4d\u7f6e\u63a5\u8fd1\u201d<\/h3>\n<p>\u8fd9\u4e00\u7bc7\u8bba\u6587\u6700\u6838\u5fc3\u3001\u4e5f\u662f\u6700\u503c\u5f97\u7406\u89e3\u7684\u601d\u60f3&#xff0c;\u53ef\u4ee5\u6982\u62ec\u4e3a&#xff1a;<\/p>\n<p>Content Similarity<br \/>\n        \u2193<br \/>\nChannel-wise Sorting<br \/>\n        \u2193<br \/>\nRelative Position<br \/>\n        \u2193<br \/>\n1D Convolution<br \/>\n        \u2193<br \/>\nContent-driven Aggregation<\/p>\n<p>\u666e\u901a convolution \u4e0d\u8ba4\u8bc6 content similarity&#xff0c;\u4f46\u662f\u5b83\u975e\u5e38\u64c5\u957f\u5229\u7528 relative position\u3002<\/p>\n<p>\u90a3\u4e48\u4f5c\u8005\u5c31\u63d0\u51fa&#xff1a;<\/p>\n<p>\u5982\u679c\u80fd\u591f\u901a\u8fc7\u6392\u5e8f&#xff0c;\u8ba9\u5185\u5bb9\u8d8a\u76f8\u4f3c\u7684 token \u5728\u65b0\u7684\u5e8f\u5217\u4e2d\u8ddd\u79bb\u8d8a\u8fd1&#xff0c;\u90a3\u4e48 convolution \u867d\u7136\u4f9d\u7136\u662f\u6309\u7167\u4f4d\u7f6e\u8fdb\u884c\u805a\u5408&#xff0c;\u4f46\u662f\u8fd9\u4e2a\u201c\u4f4d\u7f6e\u201d\u5b9e\u9645\u4e0a\u5df2\u7ecf\u5305\u542b\u4e86 content similarity\u3002<\/p>\n<p>\u4e3e\u4e00\u4e2a\u7b80\u5355\u4f8b\u5b50\u3002<\/p>\n<p>\u5047\u8bbe\u67d0\u4e2a channel \u4e2d\u7684 feature value \u4e3a&#xff1a;<\/p>\n<p>\u539f\u59cb\u987a\u5e8f&#xff1a;<\/p>\n<p>0.10    3.20    0.20    7.80    3.30    0.15<\/p>\n<p>\u4ece\u7a7a\u95f4\u4f4d\u7f6e\u6765\u770b&#xff0c;0.10 \u548c 0.15 \u8ddd\u79bb\u5f88\u8fdc&#xff0c;3.20 \u548c 3.30 \u4e5f\u5e76\u4e0d\u76f8\u90bb\u3002<\/p>\n<p>\u4f46\u662f\u8fdb\u884c\u5347\u5e8f\u6392\u5e8f\u4ee5\u540e&#xff1a;<\/p>\n<p>0.10    0.15    0.20    3.20    3.30    7.80<\/p>\n<p>\u6b64\u65f6&#xff1a;<\/p>\n<p>0.10 \u4e0e 0.15<\/p>\n<p>\u53d8\u6210\u4e86\u90bb\u5c45&#xff1b;<\/p>\n<p>3.20 \u4e0e 3.30<\/p>\n<p>\u4e5f\u53d8\u6210\u4e86\u90bb\u5c45\u3002<\/p>\n<p>\u5982\u679c\u73b0\u5728\u518d\u8fdb\u884c local 1D convolution&#xff0c;\u90a3\u4e48\u5377\u79ef\u7a97\u53e3\u5185\u90e8\u5305\u542b\u7684\u5c31\u4e0d\u518d\u53ea\u662f\u201c\u539f\u59cb\u7a7a\u95f4\u4e0a\u76f8\u90bb\u201d\u7684\u5143\u7d20&#xff0c;\u800c\u662f&#xff1a;<\/p>\n<p>feature value \u6bd4\u8f83\u76f8\u4f3c\u7684\u5143\u7d20\u3002<\/p>\n<p>\u6240\u4ee5\u6574\u4e2a\u65b9\u6cd5\u7684\u5173\u952e\u5e76\u4e0d\u662f\u8ba1\u7b97\u4e00\u4e2a\u590d\u6742\u7684 similarity matrix&#xff0c;\u800c\u662f&#xff1a;<\/p>\n<p>\u5229\u7528 sorting \u5c06 similarity \u6620\u5c04\u4e3a positional proximity\u3002<\/p>\n<p>\u8fd9\u4e5f\u662f\u8bba\u6587\u9898\u76ee Convolutional Neural Networks Driven by Content Similarity \u7684\u771f\u6b63\u542b\u4e49\u3002<\/p>\n<hr \/>\n<h3>\u4e09\u3001\u4ece Gated Convolution \u91cd\u65b0\u7406\u89e3\u5377\u79ef<\/h3>\n<h4>3.1 \u666e\u901a Gated Convolution<\/h4>\n<p>\u4f5c\u8005\u9996\u5148\u4ece\u8fd1\u5e74\u6765 CNN token mixer \u4e2d\u6bd4\u8f83\u5e38\u89c1\u7684 gated convolution \u51fa\u53d1\u3002<\/p>\n<p>\u5bf9\u4e8e\u8f93\u5165&#xff1a;<\/p>\n<p style=\"text-align:center\"><img decoding=\"async\" alt=\"X\\\\in\\\\mathbb{R}^{H\\\\times W\\\\times C}\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074948-6a7d771cb7721.png\" \/><\/p>\n<p>\u5176\u4e2d&#xff0c;H \u548c W \u4e3a feature map \u7684\u7a7a\u95f4\u5c3a\u5bf8&#xff0c;C \u4e3a channel \u6570\u91cf\u3002<\/p>\n<p>\u666e\u901a gated convolution \u53ef\u4ee5\u8868\u793a\u4e3a&#xff1a;<\/p>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"128\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074948-6a7d771cc33e1.png\" width=\"266\" \/><\/p>\n<p>\u5176\u4e2d&#xff0c;(<img decoding=\"async\" alt=\"\\\\operatorname{Conv}{pw})\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074948-6a7d771cd51b2.png\" \/> \u8868\u793a point-wise convolution&#xff0c;(<img decoding=\"async\" alt=\"\\\\operatorname{Conv}{dw}\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074948-6a7d771cdf1ae.png\" \/>) \u8868\u793a depth-wise convolution&#xff0c;(<img decoding=\"async\" alt=\"\\\\odot\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074948-6a7d771ce8d83.png\" \/>) \u8868\u793a Hadamard product\u3002<\/p>\n<p>\u8fd9\u91cc\u7684\u57fa\u672c\u8fc7\u7a0b\u53ef\u4ee5\u7406\u89e3\u4e3a&#xff1a;<\/p>\n<p>X<br \/>\n\u251c\u2500\u2500\u2192 Point-wise Conv \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2192 V<br \/>\n\u2502<br \/>\n\u2514\u2500\u2500\u2192 Point-wise Conv<br \/>\n          \u2193<br \/>\n      DWConv<br \/>\n          \u2193<br \/>\n          A<br \/>\n          \u2193<br \/>\n      V \u2299 A<br \/>\n          \u2193<br \/>\n          Z<\/p>\n<p>A \u9996\u5148\u901a\u8fc7 depth-wise convolution \u805a\u5408\u5c40\u90e8\u7a7a\u95f4\u4fe1\u606f&#xff0c;\u7136\u540e\u4f5c\u4e3a\u4e00\u4e2a\u7c7b\u4f3c attention weight \u7684 gating signal \u53bb\u8c03\u5236 V\u3002<\/p>\n<p>\u4f46\u662f\u8fd9\u79cd\u5199\u6cd5\u548c Self-Attention \u770b\u8d77\u6765\u4ecd\u7136\u5dee\u522b\u6bd4\u8f83\u5927&#xff0c;\u56e0\u4e3a Self-Attention \u901a\u5e38\u662f\u5229\u7528 Query \u548c Key \u8ba1\u7b97 attention weight&#xff0c;\u518d\u5bf9 Value \u8fdb\u884c\u52a0\u6743\u6c42\u548c\u3002<\/p>\n<p>\u56e0\u6b64\u4f5c\u8005\u8fdb\u4e00\u6b65\u6362\u4e86\u4e00\u4e2a\u89c6\u89d2\u3002<\/p>\n<hr \/>\n<h4>3.2 \u5c06 Gated Convolution \u5199\u6210\u7c7b\u4f3c Attention \u7684\u5f62\u5f0f<\/h4>\n<p>\u4f5c\u8005\u91cd\u65b0\u5b9a\u4e49&#xff1a;<\/p>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"76\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074948-6a7d771cf2760.png\" width=\"170\" \/><\/p>\n<p>\u5e76\u5c06\u8f93\u51fa\u5199\u6210&#xff1a;<\/p>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"43\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d0b1be.png\" width=\"228\" \/><\/p>\n<p>\u5176\u4e2d&#xff1a;<\/p>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"64\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d198ce.png\" width=\"373\" \/><\/p>\n<p>\u8fd9\u91cc&#xff1a;<\/p>\n<ul>\n<li>\n<p>((i,j)) \u8868\u793a\u5f53\u524d token \u7684\u7a7a\u95f4\u4f4d\u7f6e&#xff1b;<\/p>\n<\/li>\n<li>\n<p>((p,q)) \u8868\u793a\u5176\u4ed6 token \u7684\u4f4d\u7f6e&#xff1b;<\/p>\n<\/li>\n<li>\n<p>c \u8868\u793a channel&#xff1b;<\/p>\n<\/li>\n<li>\n<p>(<img decoding=\"async\" alt=\"\\\\Omega_{i,j}\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d2a208.png\" \/>) \u8868\u793a\u4ee5 ((i,j)) \u4e3a\u4e2d\u5fc3\u7684 local window&#xff1b;<\/p>\n<\/li>\n<li>\n<p>(<img decoding=\"async\" alt=\"W_{i,j,p,q,c}\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d33f47.png\" \/>) \u4e3a convolution parameter\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u4ece\u5f62\u5f0f\u4e0a\u770b&#xff0c;\u5b83\u5df2\u7ecf\u548c Attention \u6bd4\u8f83\u76f8\u4f3c&#xff1a;<\/p>\n<p>\u5f97\u5230 K<br \/>\n\u2193<br \/>\n\u751f\u6210 aggregation weight<br \/>\n\u2193<br \/>\n\u5bf9 V \u52a0\u6743<br \/>\n\u2193<br \/>\n\u5f97\u5230 Z<\/p>\n<p>\u4f46\u662f\u7ee7\u7eed\u5206\u6790\u5c31\u4f1a\u53d1\u73b0\u4e00\u4e2a\u5173\u952e\u95ee\u9898\u3002<\/p>\n<p>\u5bf9\u4e8e\u4e0d\u540c\u7684 (<img decoding=\"async\" alt=\"V_{p,q,c}\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d3ed00.png\" \/>)&#xff0c;\u5f53\u524d (<img decoding=\"async\" alt=\"K_{i,j,c}\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d49675.png\" \/>) \u662f\u5171\u4eab\u7684&#xff0c;\u56e0\u6b64\u771f\u6b63\u51b3\u5b9a\u4e0d\u540c\u4f4d\u7f6e attention weight \u5dee\u5f02\u7684\u5b9e\u9645\u4e0a\u662f&#xff1a;<\/p>\n<p style=\"text-align:center\"><img decoding=\"async\" alt=\"W_{i,j,p,q,c}\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d33f47.png\" \/><\/p>\n<p>\u800c\u4f20\u7edf convolution \u4e2d\u7684\u8fd9\u4e2a W&#xff0c;\u53ea\u8ddfrelative position\u6709\u5173&#xff0c;\u5e76\u4e0d\u8ddf<img decoding=\"async\" alt=\"K_{i,j,c}\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d49675.png\" \/>\u548c<img decoding=\"async\" alt=\"K_{p,q,c}\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d53037.png\" \/>\u4e4b\u95f4\u5177\u4f53\u7684 content similarity \u6709\u76f4\u63a5\u5173\u7cfb\u3002<\/p>\n<p>\u56e0\u6b64\u4f5c\u8005\u8ba4\u4e3a&#xff1a;<\/p>\n<p>Gated Convolution \u548c Self-Attention \u4e4b\u95f4\u771f\u6b63\u5173\u952e\u7684\u5dee\u5f02&#xff0c;\u4ecd\u7136\u5728\u4e8e convolution weight \u662f position-driven&#xff0c;\u800c Self-Attention weight \u662f content-driven\u3002<\/p>\n<hr \/>\n<h3>\u56db\u3001Convolution driven by Content Similarity<\/h3>\n<h4>4.1 \u6700\u5173\u952e\u7684\u95ee\u9898<\/h4>\n<p>\u4f5c\u8005\u63a5\u4e0b\u6765\u63d0\u51fa\u4e86\u4e00\u4e2a\u975e\u5e38\u81ea\u7136\u7684\u95ee\u9898&#xff1a;<\/p>\n<p>\u80fd\u4e0d\u80fd\u8ba9 convolution parameter \u867d\u7136\u4ecd\u7136\u6839\u636e relative position \u5de5\u4f5c&#xff0c;\u4f46\u662f relative position \u672c\u8eab\u7531 content \u51b3\u5b9a&#xff1f;<\/p>\n<p>\u4e5f\u5c31\u662f\u8bf4&#xff0c;\u5982\u679c\u6ee1\u8db3&#xff1a;<\/p>\n<p>\u5185\u5bb9\u8d8a\u76f8\u4f3c<br \/>\n    \u2193<br \/>\n\u4f4d\u7f6e\u8d8a\u63a5\u8fd1<\/p>\n<p>\u90a3\u4e48\u5c31\u53ef\u4ee5\u8fdb\u4e00\u6b65\u5f97\u5230&#xff1a;<\/p>\n<p>relative position<br \/>\n      \u2193<br \/>\nconvolution weight<br \/>\n      \u2193<br \/>\ncontent similarity<\/p>\n<p>\u4e8e\u662f\u5377\u79ef\u5c31\u53ef\u4ee5\u95f4\u63a5\u53d8\u6210 content-driven aggregation\u3002<\/p>\n<p>\u4f5c\u8005\u4e3a\u6b64\u8bbe\u8ba1\u4e86\u4e00\u6761\u65b0\u7684 parallel branch\u3002<\/p>\n<hr \/>\n<h4>4.2 Sort&#xff1a;\u9996\u5148\u5bf9 K \u8fdb\u884c Channel-wise Sorting<\/h4>\n<p>\u7b2c\u4e00\u6b65&#xff1a;<\/p>\n<h2 style=\"text-align:center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"34\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d5c84a.png\" width=\"177\" \/><\/h2>\n<p>\u8fd9\u91cc\u7684 (dw) \u8868\u793a\u6574\u4e2a sorting \u662f\u6bcf\u4e00\u4e2a channel \u72ec\u7acb\u8fdb\u884c\u7684\u3002<\/p>\n<p>\u5bf9\u4e8e\u7b2c c \u4e2a channel&#xff0c;\u5982\u679c\u91c7\u7528 ascending order&#xff1a;<\/p>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"48\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d67596.png\" width=\"482\" \/><\/p>\n<p>\u5176\u4e2d&#xff1a;<\/p>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"29\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d7555c.png\" width=\"145\" \/><\/p>\n<p>\u4e5f\u5c31\u662f\u8bf4&#xff0c;\u539f\u672c\u4e8c\u7ef4 feature map \u4e2d H\u00d7W \u4e2a\u7a7a\u95f4\u4f4d\u7f6e&#xff0c;\u88ab\u5c55\u5f00\u5e76\u6309\u7167\u8be5 channel \u4e0a\u7684 feature value \u91cd\u65b0\u6392\u5e8f\u3002<\/p>\n<p>\u8fd9\u91cc\u9700\u8981\u7279\u522b\u5f3a\u8c03\u4e00\u70b9&#xff1a;<\/p>\n<p>Ego \u7684 content similarity \u4e0d\u662f\u901a\u8fc7\u663e\u5f0f\u8ba1\u7b97\u6240\u6709 token \u7684 cosine similarity \u6216\u8005 (QK^T) \u5f97\u5230\u7684\u3002<\/p>\n<p>\u4f5c\u8005\u5e76\u6ca1\u6709\u5f62\u6210\u4e00\u4e2a<img decoding=\"async\" alt=\"N\\\\times N\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d80325.png\" \/>\u7684 similarity matrix\u3002\u800c\u662f\u5728\u6bcf\u4e2a channel \u4e2d&#xff0c;\u901a\u8fc7 feature value \u7684\u5927\u5c0f\u5173\u7cfb\u8fdb\u884c\u6392\u5e8f&#xff0c;\u4f7f\u5f97\u6570\u503c\u66f4\u52a0\u63a5\u8fd1\u7684 feature elements \u6392\u5e8f\u540e\u66f4\u52a0\u63a5\u8fd1\u3002\u56e0\u6b64\u5b83\u662f\u4e00\u79cd\u975e\u5e38\u8f7b\u91cf\u7684 implicit content similarity modeling\u3002<\/p>\n<hr \/>\n<h4>4.3 SortBy&#xff1a;V \u5fc5\u987b\u8ddf K \u4f7f\u7528\u76f8\u540c\u7684\u6392\u5217<\/h4>\n<p>\u53ea\u91cd\u65b0\u6392\u5217 K \u662f\u4e0d\u591f\u7684&#xff0c;\u56e0\u4e3a\u6700\u7ec8\u771f\u6b63\u9700\u8981\u8fdb\u884c\u4fe1\u606f\u805a\u5408\u7684\u662f V\u3002<\/p>\n<p>\u6240\u4ee5\u8fdb\u4e00\u6b65\u5b9a\u4e49&#xff1a;<\/p>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"36\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d89a30.png\" width=\"247\" \/><\/p>\n<p>\u4f8b\u5982&#xff0c;\u5982\u679c\u539f\u59cb\u4f4d\u7f6e<img decoding=\"async\" alt=\"K_{i,j,c}\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d49675.png\" \/>\u5728\u6392\u5e8f\u540e\u88ab\u653e\u5230\u4e86<img decoding=\"async\" alt=\"K&apos;_{a,c}\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d97191.png\" \/>&#xff0c;\u90a3\u4e48\u76f8\u540c\u4f4d\u7f6e\u7684<img decoding=\"async\" alt=\"V_{i,j,c}\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771da0a2b.png\" \/>\u4e5f\u5fc5\u987b\u8ddf\u7740\u653e\u5230<img decoding=\"async\" alt=\"V&apos;_{a,c}\" class=\"mathcode\" src=\"2026-08-13mourf1w3m5q.png\" \/>&#xff0c;\u56e0\u6b64 K \u548c V \u7684\u5bf9\u5e94\u5173\u7cfb\u59cb\u7ec8\u4fdd\u6301\u4e0d\u53d8\u3002\u53ef\u4ee5\u628a\u5b83\u7406\u89e3\u6210&#xff1a;<\/p>\n<p>K&#xff1a;<\/p>\n<p>\u539f\u4f4d\u7f6e&#xff1a;<br \/>\nA B C D<\/p>\n<p>\u6392\u5e8f\u540e&#xff1a;<br \/>\nC A D B<\/p>\n<p>\u90a3\u4e48 V&#xff1a;<\/p>\n<p>\u539f\u4f4d\u7f6e&#xff1a;<br \/>\na b c d<\/p>\n<p>\u4e5f\u5fc5\u987b\u53d8\u6210&#xff1a;<br \/>\nc a d b<\/p>\n<p>\u56e0\u6b64\u6392\u5e8f\u4e0d\u662f\u628a feature \u641e\u4e71&#xff0c;\u800c\u662f\u5bf9 K \u548c V \u8fdb\u884c\u4e00\u81f4\u7684 permutation\u3002<\/p>\n<hr \/>\n<h4>4.4 \u5728\u6392\u5e8f\u540e\u7684\u5e8f\u5217\u4e2d\u505a 1D Depth-wise Convolution<\/h4>\n<p>\u5b8c\u6210 sorting \u4ee5\u540e&#xff0c;\u539f\u6765\u7684\u4e8c\u7ef4<img decoding=\"async\" alt=\"H\\\\times W\" class=\"mathcode\" src=\"2026-08-13esbnnatvy11.png\" \/>\u7684feature map \u88ab\u8f6c\u6362\u6210\u957f\u5ea6<img decoding=\"async\" alt=\"N=H\\\\times W\" class=\"mathcode\" src=\"2026-08-13cyttul23fos.png\" \/>\u7684\u4e00\u7ef4\u5e8f\u5217\u3002\u4e8e\u662f\u4f5c\u8005\u5728\u65b0\u7684 content branch \u4e2d\u91c7\u75281D Depth-wise Convolution\u8fdb\u884c\u4fe1\u606f\u805a\u5408\u3002<\/p>\n<p>\u8fd9\u4e2a\u65f6\u5019\u5377\u79ef\u672c\u8eab\u4ecd\u7136\u6839\u636e relative position \u5de5\u4f5c\u3002<\/p>\n<p>\u4f46\u662f\u6ce8\u610f&#xff1a;<\/p>\n<p>\u8fd9\u4e2a relative position \u5df2\u7ecf\u4e0d\u662f\u539f\u56fe\u50cf\u4e2d\u7684\u7a7a\u95f4\u4f4d\u7f6e\u3002<\/p>\n<p>\u5b83\u5bf9\u5e94\u7684\u662f&#xff1a;<\/p>\n<p>\u6392\u5e8f\u540e\u7684 feature-value position\u3002<\/p>\n<p>\u56e0\u6b64\u4e24\u4e2a\u539f\u672c\u7a7a\u95f4\u8ddd\u79bb\u975e\u5e38\u8fdc\u7684 token&#xff0c;\u53ea\u8981\u5b83\u4eec\u5728\u67d0\u4e2a channel \u4e2d feature value \u63a5\u8fd1&#xff0c;\u5c31\u53ef\u4ee5\u5728\u6392\u5e8f\u4ee5\u540e\u53d8\u6210\u90bb\u5c45\u3002<\/p>\n<p>\u968f\u540e\u4e00\u4e2a\u5f88\u5c0f\u7684 local 1D convolution window \u5c31\u80fd\u591f\u76f4\u63a5\u805a\u5408\u5b83\u4eec\u3002<\/p>\n<p>\u8fd9\u5c31\u5f62\u6210\u4e86&#xff1a;<\/p>\n<p>Spatially far<br \/>\n&#043;<br \/>\nContent similar<\/p>\n<p>        \u2193 Sort<\/p>\n<p>Adjacent in sorted sequence<\/p>\n<p>        \u2193 1D Conv<\/p>\n<p>Direct information aggregation<\/p>\n<p>\u8fd9\u5c31\u662f Ego \u5b9e\u73b0 long-range interaction \u7684\u6838\u5fc3\u3002<\/p>\n<hr \/>\n<h4>4.5 SortBack&#xff1a;\u6700\u540e\u6062\u590d\u539f\u59cb\u7a7a\u95f4\u4f4d\u7f6e<\/h4>\n<p>\u5b8c\u6210\u4e00\u7ef4\u5e8f\u5217\u4e0a\u7684\u805a\u5408\u4e4b\u540e&#xff0c;\u8fd8\u9700\u8981\u628a\u6240\u6709 feature \u6062\u590d\u5230\u539f\u56fe\u4e2d\u7684\u4f4d\u7f6e\u3002<\/p>\n<p>\u4f5c\u8005\u5b9a\u4e49&#xff1a;<\/p>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"58\" src=\"2026-08-132b55pzgc1aa.png\" width=\"393\" \/><\/p>\n<p>\u5176\u4e2d&#xff1a;<\/p>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"57\" src=\"2026-08-130ctwfi44sy3.png\" width=\"331\" \/><\/p>\n<p>(<img decoding=\"async\" alt=\"\\\\Theta_a\" class=\"mathcode\" src=\"2026-08-13jghc4liwyk4.png\" \/>) \u8868\u793a\u6392\u5e8f\u5e8f\u5217\u4e2d\u4ee5 a \u4e3a\u4e2d\u5fc3\u7684 local window\u3002<\/p>\n<p>\u6574\u4e2a content similarity branch \u53ef\u4ee5\u603b\u7ed3\u4e3a&#xff1a;<\/p>\n<p>K<br \/>\n\u2193<br \/>\nSort<br \/>\n\u2193<br \/>\nK&#039;<\/p>\n<p>V<br \/>\n\u2193<br \/>\n\u6309\u7167 K \u7684\u6392\u5e8f\u987a\u5e8f\u91cd\u65b0\u6392\u5217<br \/>\n\u2193<br \/>\nV&#039;<\/p>\n<p>K&#039; &#043; V&#039;<br \/>\n\u2193<br \/>\n1D DWConv<br \/>\n\u2193<br \/>\nContent-based aggregation<br \/>\n\u2193<br \/>\nSortBack<br \/>\n\u2193<br \/>\n\u6062\u590d\u4e8c\u7ef4\u4f4d\u7f6e<\/p>\n<hr \/>\n<h3>\u4e94\u3001SpatialEgo&#xff1a;Spatial information \u548c Content similarity \u540c\u65f6\u4fdd\u7559<\/h3>\n<p>\u5982\u679c\u5b8c\u5168\u6309\u7167\u6392\u5e8f\u540e\u7684 feature \u8fdb\u884c\u5377\u79ef&#xff0c;\u867d\u7136\u80fd\u591f\u5f88\u597d\u5730\u805a\u5408 content-similar token&#xff0c;\u4f46\u662f\u53ef\u80fd\u635f\u5931\u539f\u59cb image structure \u4e2d\u975e\u5e38\u91cd\u8981\u7684 spatial locality\u3002<\/p>\n<p>\u56e0\u6b64\u4f5c\u8005\u6ca1\u6709\u76f4\u63a5\u5220\u9664\u4f20\u7edf\u7a7a\u95f4\u5377\u79ef&#xff0c;\u800c\u662f\u91c7\u7528\u53cc\u5206\u652f\u7ed3\u6784\u3002<\/p>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"372\" src=\"2026-08-13orixtxhp5o2.png\" width=\"368\" \/><\/p>\n<p>\u8bba\u6587 Figure 1&#xff1a;SpatialEgo token mixer \u7684\u5b8c\u6574\u7ed3\u6784\u3002<\/p>\n<p>SpatialEgo \u7684\u7b2c\u4e00\u6761 branch \u662f&#xff1a;<\/p>\n<p>V<br \/>\n\u2193<br \/>\n7\u00d77 2D Depth-wise Convolution<br \/>\n\u2193<br \/>\nSpatial aggregation<\/p>\n<p>\u5b83\u8d1f\u8d23\u5efa\u6a21\u771f\u5b9e\u4e8c\u7ef4\u7a7a\u95f4\u4e0a\u7684 local structure\u3002<\/p>\n<p>\u7b2c\u4e8c\u6761 branch \u662f&#xff1a;<\/p>\n<p>K \u2192 Sort<br \/>\n     \u2193<br \/>\nV \u2192 SortBy<br \/>\n     \u2193<br \/>\n1D Depth-wise Convolution<br \/>\n     \u2193<br \/>\nSortBack<\/p>\n<p>\u5b83\u8d1f\u8d23\u5efa\u6a21 content similarity\u3002<\/p>\n<p>\u4e24\u6761 branch \u4f7f\u7528\u76f8\u540c\u7684 K \u548c V&#xff0c;\u6700\u540e\u5c06\u4e24\u4e2a\u8f93\u51fa\u8fdb\u884c\u76f8\u52a0&#xff0c;\u7136\u540e\u518d\u901a\u8fc7 projection \u5f97\u5230\u6700\u7ec8\u8f93\u51fa\u3002<\/p>\n<p>\u56e0\u6b64 SpatialEgo \u5b9e\u9645\u4e0a\u540c\u65f6\u62e5\u6709&#xff1a;<\/p>\n<p>Spatial relation<br \/>\n&#043;<br \/>\nContent relation<\/p>\n<p>\u4f5c\u8005\u5e76\u4e0d\u662f\u7528 content similarity \u66ff\u4ee3 spatial information&#xff0c;\u800c\u662f\u8ba9\u4e24\u79cd\u4fe1\u606f\u4e92\u8865\u3002<\/p>\n<p>\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48\u8fd9\u4e2a\u6a21\u5757\u53eb&#xff1a;<\/p>\n<p>SpatialEgo<\/p>\n<p>\u800c\u4e0d\u4ec5\u4ec5\u662f\u4e00\u4e2a sorting convolution\u3002<\/p>\n<hr \/>\n<h3>\u516d\u3001\u4e3a\u4ec0\u4e48\u5c40\u90e8\u5377\u79ef\u53ef\u4ee5\u4ea7\u751f\u201c\u9690\u5f0f\u5168\u5c40\u611f\u53d7\u91ce\u201d&#xff1f;<\/h3>\n<p>\u8fd9\u662f Ego \u4e0e\u666e\u901a local convolution \u6700\u5927\u7684\u5dee\u5f02\u4e4b\u4e00\u3002<\/p>\n<p>\u5047\u8bbe\u56fe\u50cf\u5de6\u4e0a\u89d2\u5b58\u5728\u4e00\u4e2a token A&#xff0c;\u53f3\u4e0b\u89d2\u5b58\u5728\u4e00\u4e2a token B\u3002<\/p>\n<p>\u5728\u539f\u59cb\u4e8c\u7ef4\u7a7a\u95f4\u91cc&#xff1a;<\/p>\n<p>A &#8212;&#8212;&#8212;&#8212;&#8212;&#8212;&#8212;&#8212;&#8212;&#8212;- B<\/p>\n<p>\u4e24\u8005\u8ddd\u79bb\u5f88\u8fdc&#xff0c;\u4e00\u4e2a 7\u00d77 convolution \u663e\u7136\u65e0\u6cd5\u76f4\u63a5\u628a B \u7684\u4fe1\u606f\u9001\u5230 A\u3002<\/p>\n<p>\u4f46\u662f\u5982\u679c&#xff1a;<\/p>\n<p>feature(A) \u2248 feature(B)<\/p>\n<p>\u7ecf\u8fc7 sorting \u4ee5\u540e&#xff1a;<\/p>\n<p>A B<\/p>\n<p>\u4e24\u8005\u53ef\u80fd\u76f4\u63a5\u53d8\u6210\u90bb\u5c45\u3002<\/p>\n<p>\u90a3\u4e48\u5373\u4f7f\u4f7f\u7528\u975e\u5e38\u5c0f\u7684\u4e00\u7ef4 convolution window&#xff0c;\u4e5f\u53ef\u4ee5\u8ba9 A \u805a\u5408 B \u7684\u4fe1\u606f\u3002<\/p>\n<p>\u56e0\u6b64&#xff1a;<\/p>\n<p>\u5377\u79ef\u7a97\u53e3\u867d\u7136\u662f local \u7684&#xff0c;\u4f46\u7531\u4e8e\u6392\u5e8f\u64cd\u4f5c\u672c\u8eab\u5177\u6709 global rearrangement ability&#xff0c;\u6240\u4ee5\u5b83\u5bf9\u5e94\u7684\u539f\u59cb\u56fe\u50cf receptive field \u53ef\u4ee5\u662f global \u7684\u3002<\/p>\n<hr \/>\n<h4>6.1 Window Size \u7684\u8bbe\u8ba1<\/h4>\n<p>\u4f5c\u8005\u5b9a\u4e49\u4e00\u7ef4 sliding window&#xff1a;<\/p>\n<p style=\"text-align:center\"><img decoding=\"async\" alt=\"2\\\\alpha\\\\lfloor\\\\ln N\\\\rfloor+1\" class=\"mathcode\" src=\"2026-08-13aad253hlrvo.png\" \/><\/p>\n<p>\u5176\u4e2d&#xff1a;<\/p>\n<p style=\"text-align:center\"><img decoding=\"async\" alt=\"N=H\\\\times W\" class=\"mathcode\" src=\"2026-08-13cyttul23fos.png\" \/><\/p>\n<p>(<img decoding=\"async\" alt=\"\\\\alpha\" class=\"mathcode\" src=\"2026-08-13uivfi4wp2t4.png\" \/>) \u4e3a\u6b63\u6574\u6570\u3002\u8fd9\u79cd\u8bbe\u8ba1\u6709\u4e24\u4e2a\u539f\u56e0\u3002\u7b2c\u4e00&#xff0c;\u968f\u7740 N \u589e\u52a0&#xff0c;\u4e00\u4e2a feature element \u9644\u8fd1\u53ef\u80fd\u51fa\u73b0\u66f4\u591a\u76f8\u4f3c\u5143\u7d20&#xff0c;\u56e0\u6b64 window size \u4e5f\u5e94\u8be5\u9002\u5f53\u589e\u52a0\u3002\u7b2c\u4e8c&#xff0c;\u5982\u679c\u76f4\u63a5\u91c7\u7528 global convolution&#xff0c;2N-1&#xff0c;\u8ba1\u7b97\u6210\u672c\u4f1a\u968f\u8f93\u5165\u89c4\u6a21\u660e\u663e\u63d0\u9ad8&#xff0c;\u800c<img decoding=\"async\" alt=\"2\\\\alpha\\\\lfloor\\\\ln N\\\\rfloor+1\" class=\"mathcode\" src=\"2026-08-13aad253hlrvo.png\" \/>\u80fd\u591f\u4f7f\u590d\u6742\u5ea6\u4fdd\u6301\u5728<img decoding=\"async\" alt=\"O(N\\\\log N)\" class=\"mathcode\" src=\"2026-08-13tuiyrwwdofh.png\" \/>&#xff0c;\u4e0d\u9700\u8981\u4f9d\u8d56 FFT\u3002\u4f5c\u8005\u7ecf\u8fc7\u7b80\u5355\u641c\u7d22\u4ee5\u540e\u53d1\u73b0<img decoding=\"async\" alt=\"\\\\alpha=12\" class=\"mathcode\" src=\"2026-08-13w4tybb40ye1.png\" \/>\u5df2\u7ecf\u53ef\u4ee5\u8fbe\u5230 global convolution \u76f8\u540c\u7684\u6027\u80fd\u3002\u56e0\u6b64\u6700\u7ec8\u9ed8\u8ba4<img decoding=\"async\" alt=\"24\\\\lfloor\\\\ln N\\\\rfloor+1\" class=\"mathcode\" src=\"2026-08-1340opzx1gqxz.png\" \/><\/p>\n<hr \/>\n<h3>\u4e03\u3001Convolution Weight&#xff1a;\u8ddd\u79bb\u8d8a\u8fd1&#xff0c;\u6743\u91cd\u8d8a\u5927<\/h3>\n<p>\u7531\u4e8e\u6392\u5e8f\u4ee5\u540e&#xff1a;<\/p>\n<p>\u8ddd\u79bb\u8d8a\u8fd1<br \/>\n\u2248<br \/>\nfeature value \u8d8a\u63a5\u8fd1<br \/>\n\u2248<br \/>\ncontent \u8d8a\u76f8\u4f3c<\/p>\n<p>\u4f5c\u8005\u81ea\u7136\u91c7\u7528\u4e86&#xff1a;<\/p>\n<p>weight decays with distance<\/p>\n<p>\u7684\u6743\u91cd\u8bbe\u8ba1\u3002<\/p>\n<p>\u4f5c\u8005\u6d4b\u8bd5\u4e86 PyTorch \u4e2d\u4e24\u4e2a\u975e\u5e38\u7b80\u5355\u7684\u51fd\u6570&#xff1a;<\/p>\n<p>linspace<\/p>\n<p>logspace<\/p>\n<p>\u7ed3\u679c\u53d1\u73b0\u4e8c\u8005\u8868\u73b0\u51e0\u4e4e\u76f8\u540c\u3002<\/p>\n<p>\u5728 Ego-T \u4e0a\u5206\u522b\u53ef\u4ee5\u53d6\u5f97\u5927\u7ea6&#xff1a;<\/p>\n<p>83.9%<\/p>\n<p>84.0%<\/p>\n<p>\u7684 Top-1 accuracy\u3002<\/p>\n<p>\u6700\u7ec8\u4f5c\u8005\u4f7f\u7528 logspace \u751f\u6210\u5377\u79ef\u53c2\u6570\u3002<\/p>\n<p>\u5176\u5f62\u5f0f\u53ef\u4ee5\u5199\u4e3a&#xff1a;<\/p>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"98\" src=\"2026-08-13h43tcqumpga.png\" width=\"362\" \/><\/p>\n<p>\u5176\u4e2d&#xff1a;<\/p>\n<p>base&#061;10<\/p>\n<p>\u5e76\u4e14&#xff1a;<\/p>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"26\" src=\"2026-08-131ybtapsuymz.png\" width=\"171\" \/><\/p>\n<p>\u53ef\u4ee5\u770b\u5230&#xff0c;\u5f53&#xff1a;|b-a|\u8d8a\u5c0f\u65f6&#xff0c;\u4e24\u4e2a feature \u5728\u6392\u5e8f\u5e8f\u5217\u4e2d\u7684\u8ddd\u79bb\u8d8a\u8fd1&#xff0c;\u5bf9\u5e94\u7684 weight \u8d8a\u5927\u3002<\/p>\n<p>\u4f5c\u8005\u7279\u522b\u6307\u51fa&#xff0c;\u8fd9\u4e9b weight&#xff1a;<\/p>\n<p>\u53ea\u4e0e relative position \u6709\u5173&#xff0c;\u800c\u4e0e\u5177\u4f53 channel \u65e0\u5173\u3002<\/p>\n<p>\u56e0\u6b64\u53ea\u9700\u8981\u751f\u6210\u4e00\u6b21&#xff0c;\u5c31\u53ef\u4ee5\u5728\u4e0d\u540c\u4f4d\u7f6e\u4e0e channel \u4e4b\u95f4\u5171\u4eab\u3002<\/p>\n<p>\u6362\u53e5\u8bdd\u8bf4&#xff0c;Ego \u4e2d\u65b0\u589e\u7684 content-driven mechanism \u5e76\u4e0d\u9700\u8981\u518d\u989d\u5916\u8bad\u7ec3\u4e00\u4e2a\u590d\u6742 attention network \u6765\u9884\u6d4b similarity weight&#xff0c;\u800c\u662f&#xff1a;<\/p>\n<p>Sorting<br \/>\n&#043;<br \/>\nFixed distance-decaying weights<br \/>\n&#043;<br \/>\nDepth-wise convolution<\/p>\n<p>\u5b8c\u6210\u5185\u5bb9\u9a71\u52a8\u7684\u4fe1\u606f\u805a\u5408\u3002<\/p>\n<hr \/>\n<h3>\u516b\u3001Ego \u7f51\u7edc\u6574\u4f53\u7ed3\u6784<\/h3>\n<p>SpatialEgo \u53ea\u662f token mixer&#xff0c;\u4f5c\u8005\u6700\u7ec8\u5728\u5176\u57fa\u7840\u4e0a\u6784\u9020\u4e86\u5b8c\u6574\u7684 CNN backbone&#xff1a;Ego\u3002<\/p>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"435\" src=\"2026-08-13bhh45hvrz35.png\" width=\"407\" \/><\/p>\n<p>\u8bba\u6587 Figure 2&#xff1a;Ego \u7684\u6574\u4f53\u56db\u9636\u6bb5\u7f51\u7edc\u7ed3\u6784\u4ee5\u53ca Ego Block\u3002<\/p>\n<p>\u6574\u4f53\u7f51\u7edc\u91c7\u7528\u7c7b\u4f3c ConvNeXt \u548c Swin Transformer \u7684 hierarchical meta-architecture\u3002<\/p>\n<p>\u5b83\u7684 macro architecture \u4e0e\u5f88\u591a\u73b0\u4ee3 CNN \/ Transformer backbone \u5f88\u63a5\u8fd1&#xff0c;\u771f\u6b63\u4e0d\u540c\u7684\u90e8\u5206\u4e3b\u8981\u96c6\u4e2d\u5728&#xff1a;<\/p>\n<p>SpatialEgo Token Mixer<\/p>\n<p>\u8fd9\u4e00\u70b9\u4e5f\u4f7f\u8fd9\u7bc7\u8bba\u6587\u7684\u6539\u52a8\u6bd4\u8f83\u5e72\u51c0&#xff1a;\u5b83\u5e76\u6ca1\u6709\u628a\u6027\u80fd\u63d0\u5347\u5efa\u7acb\u5728\u4e00\u4e2a\u975e\u5e38\u590d\u6742\u7684\u65b0 backbone \u4e0a&#xff0c;\u800c\u662f\u91cd\u70b9\u9a8c\u8bc1\u65b0\u7684 token mixing mechanism\u3002<\/p>\n<hr \/>\n<h4>8.1 Ego-T\u3001Ego-S \u548c Ego-B<\/h4>\n<p>\u4f5c\u8005\u6784\u9020\u4e09\u4e2a\u4e0d\u540c\u89c4\u6a21\u7684\u7248\u672c\u3002<\/p>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"119\" src=\"2026-08-13upauc1ctajk.png\" width=\"449\" \/><\/p>\n<p>\u8bba\u6587 Table 1&#xff1a;Ego-T\u3001Ego-S \u548c Ego-B \u7684 channel configuration \u4e0e\u6bcf\u4e2a stage \u7684 block \u6570\u91cf\u3002<\/p>\n<p>\u4ece T \u2192 S \u4e3b\u8981\u589e\u52a0 block depth&#xff0c;\u800c\u4ece S \u2192 B \u53c8\u8fdb\u4e00\u6b65\u589e\u52a0 channel width\u3002<\/p>\n<hr \/>\n<h3>\u4e5d\u3001ImageNet-1K \u56fe\u50cf\u5206\u7c7b\u5b9e\u9a8c<\/h3>\n<p>\u4f5c\u8005\u9996\u5148\u5728 ImageNet-1K \u4e0a\u6d4b\u8bd5 Ego\u3002<\/p>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"318\" src=\"2026-08-13rhyp0dp2hka.png\" width=\"402\" \/><\/p>\n<p>\u8bba\u6587 Table 2&#xff1a;ImageNet-1K \u4e0a\u8bad\u7ec3 Ego-T\/S\/B \u4f7f\u7528\u7684\u5b8c\u6574\u8d85\u53c2\u6570\u914d\u7f6e\u3002<\/p>\n<hr \/>\n<h4>9.1 ImageNet-1K \u7ed3\u679c<\/h4>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"577\" src=\"2026-08-131wwpvhby0wd.png\" width=\"349\" \/><\/p>\n<p>\u8bba\u6587 Table 3&#xff1a;Ego \u4e0e CNN\u3001Transformer \u548c Hybrid models \u5728 ImageNet-1K \u4e0a\u7684 Params\u3001FLOPs \u4e0e Top-1 Accuracy \u5bf9\u6bd4\u3002<\/p>\n<hr \/>\n<h4>9.2 Ego-S \u7684\u53c2\u6570\u6548\u7387\u975e\u5e38\u7a81\u51fa<\/h4>\n<p>Ego \u7684 content-similarity-driven convolution \u5e76\u4e0d\u662f\u53ea\u5bf9 Tiny model \u6709\u6548&#xff0c;\u968f\u7740\u6a21\u578b scale \u589e\u52a0\u4ecd\u7136\u80fd\u591f\u6301\u7eed\u5f97\u5230\u8f83\u597d\u7684 performance\u3002<\/p>\n<hr \/>\n<h3>\u5341\u3001ADE20K Semantic Segmentation<\/h3>\n<p>\u5206\u7c7b\u51c6\u786e\u7387\u53ea\u80fd\u8bf4\u660e backbone \u7684\u5168\u5c40\u8bc6\u522b\u80fd\u529b&#xff0c;\u56e0\u6b64\u4f5c\u8005\u7ee7\u7eed\u5c06 ImageNet-1K pretrained Ego \u7528\u4f5c downstream dense prediction backbone\u3002<\/p>\n<p>Semantic segmentation \u4f7f\u7528&#xff1a;<\/p>\n<p>ADE20K<\/p>\n<p>ADE20K \u5305\u542b&#xff1a;<\/p>\n<p>20K training images<\/p>\n<p>2K validation images<\/p>\n<p>150 semantic classes<\/p>\n<p>\u4f5c\u8005\u91c7\u7528&#xff1a;<\/p>\n<p>UperNet<\/p>\n<p>\u4f5c\u4e3a segmentation framework\u3002<\/p>\n<p>\u8f93\u5165\u56fe\u50cf random crop&#xff1a;<\/p>\n<p>512\u00d7512<\/p>\n<p>\u8bad\u7ec3&#xff1a;<\/p>\n<p>160K iterations<\/p>\n<p>optimizer&#xff1a;<\/p>\n<p>AdamW<\/p>\n<p>batch size&#xff1a;<\/p>\n<p>16<\/p>\n<hr \/>\n<h4>10.1 ADE20K \u7ed3\u679c<\/h4>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"532\" src=\"2026-08-13ti03bmzsxqb.png\" width=\"328\" \/><\/p>\n<p>\u8bba\u6587 Table 4&#xff1a;\u4e0d\u540c backbone \u5728 UperNet &#043; ADE20K \u4e0a\u7684 Params\u3001FLOPs \u4e0e mIoU\u3002<\/p>\n<hr \/>\n<h4>10.2 \u4e0e\u5f3a\u8c03 Global Modeling \u7684\u6a21\u578b\u6bd4\u8f83<\/h4>\n<p>\u8bba\u6587\u8fd8\u7279\u522b\u6bd4\u8f83\u4e86&#xff1a;<\/p>\n<p>VMamba<\/p>\n<p>OverLoCK<\/p>\n<p>VMamba \u5f3a\u8c03 global modeling&#xff0c;\u800c OverLoCK \u5f3a\u8c03 large receptive field\u3002<\/p>\n<p>Tiny scale&#xff1a;<\/p>\n<p>VMamba-T&#xff1a;48.8<\/p>\n<p>Ego-T&#xff1a;49.0<\/p>\n<p>Small scale&#xff1a;<\/p>\n<p>OverLock-T&#xff1a;50.8<\/p>\n<p>Ego-S&#xff1a;51.0<\/p>\n<p>Base scale&#xff1a;<\/p>\n<p>OverLock-S&#xff1a;51.9<\/p>\n<p>Ego-B&#xff1a;52.3<\/p>\n<p>\u6b64\u5916&#xff1a;<\/p>\n<p>VAN-B4&#xff1a;52.2<\/p>\n<p>Ego-B&#xff1a;52.3<\/p>\n<p>\u8bf4\u660e SpatialEgo \u4e0d\u53ea\u662f ImageNet classification \u6709\u63d0\u5347&#xff0c;\u5bf9\u4e8e\u9700\u8981 dense spatial prediction \u7684 segmentation \u540c\u6837\u5177\u6709\u8f83\u597d\u7684 transfer ability\u3002<\/p>\n<hr \/>\n<h3>\u5341\u4e00\u3001COCO Object Detection \u548c Instance Segmentation<\/h3>\n<p>\u4f5c\u8005\u8fdb\u4e00\u6b65\u5728COCO 2017\u4e0a\u6d4b\u8bd5\u6a21\u578b\u3002<\/p>\n<p>\u6570\u636e\u96c6&#xff1a;<\/p>\n<p>118K training images<\/p>\n<p>5K validation images<\/p>\n<p>\u68c0\u6d4b\u6846\u67b6\u91c7\u7528Cascade Mask R-CNN&#xff0c;Ego backbone \u9996\u5148\u5728 ImageNet-1K \u4e0a\u9884\u8bad\u7ec3\u3002<\/p>\n<p>\u8bad\u7ec3&#xff1a;<\/p>\n<p>36 epochs<\/p>\n<p>3\u00d7 training schedule<\/p>\n<p>multi-scale training<\/p>\n<p>optimizer&#xff1a;<\/p>\n<p>AdamW<\/p>\n<p>\u8fd9\u91cc\u8bba\u6587\u7279\u522b\u62a5\u544a\u4e86\u4e00\u4e2a\u5de5\u7a0b\u95ee\u9898\u3002<\/p>\n<p>\u7531\u4e8e COCO \u56fe\u50cf\u5206\u8fa8\u7387\u8f83\u9ad8&#xff0c;\u540c\u65f6\u4f7f\u7528 multi-scale training&#xff0c;\u5f53 Ego \u4f5c\u4e3a backbone \u65f6&#xff0c;\u5728 RTX 3090 \u4e0a\u4f7f\u7528 MMDetection \u9ed8\u8ba4&#xff1a;<\/p>\n<p>batch size&#061;16<\/p>\n<p>\u4f1a\u53d1\u751f&#xff1a;<\/p>\n<p>Memory Overflow<\/p>\n<p>\u56e0\u6b64\u4f5c\u8005\u4e0d\u5f97\u4e0d\u5c06&#xff1a;<\/p>\n<p>batch size&#061;16<\/p>\n<p>\u964d\u4f4e\u4e3a&#xff1a;<\/p>\n<p>batch size&#061;8<\/p>\n<p>\u5e76\u540c\u6b65\u8c03\u6574 learning rate \u7b49\u8d85\u53c2\u6570\u3002<\/p>\n<p>\u4f5c\u8005\u8ba4\u4e3a\u8fd9\u53ef\u80fd\u5bf9\u6a21\u578b\u6700\u7ec8\u6027\u80fd\u4ea7\u751f\u4e00\u5b9a\u8d1f\u9762\u5f71\u54cd&#xff0c;\u540c\u65f6\u5b9e\u9a8c\u8868\u660e Ego \u5bf9 training configuration \u7684\u53d8\u5316\u5b58\u5728\u4e00\u5b9a\u654f\u611f\u6027&#xff0c;\u8fd9\u4e00\u70b9\u5728 Ego-T \u4e0a\u5c24\u5176\u660e\u663e\u3002<\/p>\n<hr \/>\n<h4>11.1 COCO \u7ed3\u679c<\/h4>\n<p class=\"img-center\"><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"505\" src=\"2026-08-13twtyx3pzamw.png\" width=\"349\" \/><\/p>\n<p>\u8bba\u6587 Table 5&#xff1a;Cascade Mask R-CNN &#043; COCO \u4e0a\u4e0d\u540c backbone \u7684 Params\u3001FLOPs\u3001bbox AP \u548c mask AP\u3002<\/p>\n<p>\u9700\u8981\u6ce8\u610f\u7684\u662f&#xff0c;Ego-B \u7684 mask AP&#061;45.8&#xff0c;\u5e76\u4e0d\u662f\u8868\u683c\u4e2d\u7edd\u5bf9\u6700\u9ad8&#xff0c;\u4f8b\u5982 MogaNet-B \u4e3a 46.0&#xff0c;UniRepLKNet-S \u4e3a 45.9\u3002\u56e0\u6b64\u8fd9\u90e8\u5206\u66f4\u51c6\u786e\u7684\u7ed3\u8bba\u5e94\u8be5\u662f&#xff1a;<\/p>\n<p>Ego \u5728 detection \u548c instance segmentation \u4e2d\u53d6\u5f97\u4e86\u5f88\u6709\u7ade\u4e89\u529b\u7684\u8868\u73b0&#xff0c;\u5c24\u5176\u968f\u7740\u6a21\u578b\u89c4\u6a21\u589e\u5927&#xff0c;bbox detection performance \u7684\u4f18\u52bf\u66f4\u52a0\u660e\u663e&#xff0c;\u4f46\u5e76\u4e0d\u662f\u6240\u6709 COCO \u6307\u6807\u90fd\u53d6\u5f97\u7edd\u5bf9\u6700\u4f18\u3002<\/p>\n<hr \/>\n<h3>\u5341\u4e8c\u3001\u6d88\u878d\u5b9e\u9a8c&#xff1a;Ego \u7684\u63d0\u5347\u5230\u5e95\u6765\u81ea\u54ea\u91cc&#xff1f;<\/h3>\n<p>\u4f5c\u8005\u5728 ImageNet-1K \u4e0a\u4ee5 Ego-T \u4e3a\u57fa\u7840\u8fdb\u884c\u6d88\u878d\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"558\" src=\"2026-08-13a4eq3ckqgv1.png\" width=\"1494\" \/><\/p>\n<p>\u8bba\u6587 Table 6&#xff1a;Ego-T \u5728 ImageNet-1K \u4e0a\u5173\u4e8e Window Size\u3001Weight Generation Function \u548c Token Mixer \u7684\u6d88\u878d\u5b9e\u9a8c\u3002<\/p>\n<p>Baseline Ego-T&#xff1a;<\/p>\n<p>Params&#061;27M<\/p>\n<p>FLOPs&#061;3.8G<\/p>\n<p>Top-1&#061;84.0%<\/p>\n<hr \/>\n<h4>12.1 Window Size<\/h4>\n<p>\u9ed8\u8ba4&#xff1a;<img decoding=\"async\" alt=\"24\\\\lfloor\\\\ln N\\\\rfloor+1\" class=\"mathcode\" src=\"2026-08-1340opzx1gqxz.png\" \/><\/p>\n<p>Top-1&#xff1a;<\/p>\n<p>84.0%<\/p>\n<p>\u51cf\u5c11\u5230&#xff1a;<img decoding=\"async\" alt=\"6\\\\lfloor\\\\ln N\\\\rfloor+1\" class=\"mathcode\" src=\"2026-08-13z30wvnxyeiy.png\" \/><\/p>\n<p>\u7ed3\u679c&#xff1a;<\/p>\n<p>83.7%<\/p>\n<p>\u51cf\u5c11\u5230&#xff1a;<img decoding=\"async\" alt=\"12\\\\lfloor\\\\ln N\\\\rfloor+1\" class=\"mathcode\" src=\"2026-08-1343d24efurq0.png\" \/><\/p>\n<p>\u540c\u6837&#xff1a;<\/p>\n<p>83.7%<\/p>\n<p>\u5982\u679c\u6539\u6210 global 1D convolution&#xff1a;2N-1<\/p>\n<p>\u7ed3\u679c&#xff1a;<\/p>\n<p>84.0%<\/p>\n<p>\u4e0e\u9ed8\u8ba4\u7a97\u53e3\u5b8c\u5168\u76f8\u540c\u3002<\/p>\n<p>\u4f46\u662f FLOPs&#xff1a;<\/p>\n<p>3.8G \u2192 3.9G<\/p>\n<p>\u8fd9\u7ec4\u5b9e\u9a8c\u975e\u5e38\u91cd\u8981&#xff0c;\u56e0\u4e3a\u5b83\u76f4\u63a5\u9a8c\u8bc1\u4e86\u8bba\u6587\u7684\u6838\u5fc3\u5047\u8bbe&#xff1a;<\/p>\n<p>\u6392\u5e8f\u80fd\u591f\u628a\u7a7a\u95f4\u4e0a\u5f88\u8fdc\u4f46\u5185\u5bb9\u76f8\u4f3c\u7684\u5143\u7d20\u79fb\u52a8\u5230\u9644\u8fd1&#xff0c;\u56e0\u6b64\u4e0d\u9700\u8981\u771f\u6b63\u505a\u5168\u5c40 convolution&#xff0c;\u4e00\u4e2a (24<img decoding=\"async\" alt=\"\\\\lfloor\\\\ln N\\\\rfloor\" class=\"mathcode\" src=\"2026-08-13b0xfwm0hnmn.png\" \/>&#043;1) \u7684\u5c40\u90e8\u7a97\u53e3\u5df2\u7ecf\u80fd\u591f\u83b7\u5f97\u4e0e global convolution \u4e00\u6837\u7684\u5206\u7c7b\u51c6\u786e\u7387\u3002<\/p>\n<hr \/>\n<h4>12.2 logspace \u8fd8\u662f linspace&#xff1f;<\/h4>\n<p>\u4f5c\u8005\u8fdb\u4e00\u6b65\u5c06\u9ed8\u8ba4 logspace \u6362\u6210 linspace\u3002<\/p>\n<p>\u5bf9\u5e94&#xff1a;<img decoding=\"async\" alt=\"6\\\\lfloor\\\\ln N\\\\rfloor+1\" class=\"mathcode\" src=\"2026-08-13z30wvnxyeiy.png\" \/>\u00a0Top-1&#xff1a;<\/p>\n<p>83.8%<\/p>\n<p><img decoding=\"async\" alt=\"12\\\\lfloor\\\\ln N\\\\rfloor+1\" class=\"mathcode\" src=\"2026-08-1343d24efurq0.png\" \/>\u00a0Top-1&#xff1a;<\/p>\n<p>83.8%<\/p>\n<p>\u9ed8\u8ba4&#xff1a;<img decoding=\"async\" alt=\"24\\\\lfloor\\\\ln N\\\\rfloor+1\" class=\"mathcode\" src=\"2026-08-1340opzx1gqxz.png\" \/><\/p>\n<p>\u5f97\u5230&#xff1a;<\/p>\n<p>83.9%<\/p>\n<p>global&#xff1a;2N-1<\/p>\n<p>\u540c\u6837&#xff1a;<\/p>\n<p>83.9%<\/p>\n<p>\u4e0e logspace \u7684\u5dee\u8ddd\u59cb\u7ec8&#xff1a;<\/p>\n<p>\u22640.1%<\/p>\n<p>\u8bf4\u660e Ego \u7684\u6548\u679c\u5e76\u4e0d\u4f9d\u8d56\u4e00\u4e2a\u975e\u5e38\u590d\u6742\u6216\u8005\u7279\u522b\u7cbe\u7ec6\u7684 weight generation function\u3002<\/p>\n<p>\u771f\u6b63\u5173\u952e\u7684\u662f&#xff1a;<\/p>\n<p>\u6392\u5e8f\u4ee5\u540e&#xff0c;\u4f7f\u7528\u4e00\u4e2a\u968f\u8ddd\u79bb\u8870\u51cf\u7684 convolution weight\u3002<\/p>\n<p>\u800c\u4e0d\u662f\u4e00\u5b9a\u5fc5\u987b\u4f7f\u7528 logspace\u3002<\/p>\n<hr \/>\n<h4>12.3 \u53bb\u6389 Content Similarity Branch \u4f1a\u600e\u6837&#xff1f;<\/h4>\n<p>\u8fd9\u662f\u6700\u5173\u952e\u7684\u4e00\u7ec4\u6d88\u878d\u3002<\/p>\n<p>Baseline&#xff1a;<\/p>\n<p>[SpatialEgo,<br \/>\n SpatialEgo,<br \/>\n SpatialEgo,<br \/>\n SpatialEgo]<\/p>\n<p>\u7ed3\u679c&#xff1a;<\/p>\n<p>84.0%<\/p>\n<p>\u5982\u679c\u56db\u4e2a stage \u5168\u90e8\u6539\u6210&#xff1a;<\/p>\n<p>[GatedConv,<br \/>\n GatedConv,<br \/>\n GatedConv,<br \/>\n GatedConv]<\/p>\n<p>\u7ed3\u679c&#xff1a;<\/p>\n<p>83.5%<\/p>\n<p>\u4e0b\u964d&#xff1a;<\/p>\n<p>0.5%<\/p>\n<p>\u540c\u65f6&#xff1a;<\/p>\n<p>Params&#061;27M<\/p>\n<p>\u4fdd\u6301\u4e0d\u53d8&#xff1b;<\/p>\n<p>FLOPs&#061;3.8G<\/p>\n<p>\u4e5f\u4fdd\u6301\u4e0d\u53d8\u3002<\/p>\n<p>\u8fd9\u610f\u5473\u7740\u6027\u80fd\u5dee\u5f02\u5e76\u4e0d\u662f\u7b80\u5355\u6765\u81ea&#xff1a;<\/p>\n<p>\u66f4\u591a\u53c2\u6570<\/p>\n<p>\u6216\u8005&#xff1a;<\/p>\n<p>\u66f4\u9ad8 FLOPs<\/p>\n<p>\u800c\u662f\u4e0e\u4f5c\u8005\u63d0\u51fa\u7684 content-similarity-driven 1D branch \u76f4\u63a5\u76f8\u5173\u3002<\/p>\n<p>\u4ece\u8fd9\u4e2a\u89d2\u5ea6\u770b&#xff0c;\u8fd9\u4e2a 0.5% \u7684\u63d0\u5347\u5176\u5b9e\u6bd4\u5355\u72ec\u770b\u6570\u503c\u66f4\u52a0\u6709\u610f\u4e49&#xff0c;\u56e0\u4e3a&#xff1a;<\/p>\n<p>\u5b83\u662f\u5728\u51e0\u4e4e\u4e0d\u6539\u53d8\u6a21\u578b\u53c2\u6570\u91cf\u548c\u8ba1\u7b97\u91cf\u7684\u60c5\u51b5\u4e0b\u83b7\u5f97\u7684\u3002<\/p>\n<hr \/>\n<h3>\u5341\u4e09\u3001Visualization&#xff1a;Ego \u662f\u5426\u771f\u7684\u83b7\u5f97\u4e86\u66f4\u5927\u8303\u56f4\u7684\u4fe1\u606f\u4ea4\u4e92&#xff1f;<\/h3>\n<p>\u4f5c\u8005\u6700\u540e\u5229\u7528 influence map \u5bf9 effective receptive field \u8fdb\u884c\u53ef\u89c6\u5316\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" alt=\"\" height=\"621\" src=\"2026-08-130kdz00qdqpn.png\" width=\"855\" \/><\/p>\n<p>\u8bba\u6587 Figure 3&#xff1a;Ego \u4e0e\u53bb\u6389\u4e00\u7ef4 content branch \u540e\u7684 GatedConv \u5728 Stage 1\u3001Stage 2 \u548c Stage 3 \u4e2d\u7684 influence maps\u3002\u4e0a\u65b9\u4e3a Ego&#xff0c;\u4e0b\u65b9\u4e3a\u666e\u901a GatedConv\u3002<\/p>\n<p>\u4ece Figure 3 \u53ef\u4ee5\u660e\u663e\u770b\u5230&#xff0c;\u5728\u7f51\u7edc\u524d\u51e0\u4e2a stage \u4e2d&#xff1a;<\/p>\n<p>Ego \u7684 influence region \u66f4\u5e7f\u3002<\/p>\n<p>\u666e\u901a GatedConv \u66f4\u591a\u96c6\u4e2d\u5728 anchor point \u5468\u56f4\u7684\u5c40\u90e8\u533a\u57df\u3002<\/p>\n<p>\u800c Ego \u7684\u5f71\u54cd\u533a\u57df\u53ef\u4ee5\u8de8\u8d8a\u6bd4\u8f83\u8fdc\u7684\u7a7a\u95f4\u4f4d\u7f6e\u3002<\/p>\n<p>\u8fd9\u4e2a\u7ed3\u679c\u6070\u597d\u7b26\u5408 SpatialEgo \u7684\u8bbe\u8ba1\u903b\u8f91&#xff1a;<\/p>\n<p>\u4e24\u4e2a token \u7a7a\u95f4\u8ddd\u79bb\u5f88\u8fdc<br \/>\n        \u2193<br \/>\nfeature value \u76f8\u4f3c<br \/>\n        \u2193<br \/>\n\u7ecf\u8fc7 channel-wise sorting<br \/>\n        \u2193<br \/>\n\u5728\u4e00\u7ef4\u5e8f\u5217\u4e0a\u53d8\u5f97\u63a5\u8fd1<br \/>\n        \u2193<br \/>\n1D DWConv \u5efa\u7acb\u4fe1\u606f\u4ea4\u4e92<\/p>\n<p>\u56e0\u6b64\u867d\u7136\u5b9e\u9645\u8fdb\u884c\u7684\u4ecd\u7136\u662f local 1D convolution&#xff0c;\u4f46\u662f\u5176\u5bf9\u5e94\u5230\u539f\u59cb image plane \u4e0a\u65f6&#xff0c;\u53ef\u80fd\u5f62\u6210 long-range interaction\u3002<\/p>\n<p>\u8fd9\u4e5f\u662f\u8bba\u6587\u6240\u8bf4&#xff1a;<\/p>\n<p>Local window for global context<\/p>\n<p>\u7684\u76f4\u89c2\u8bc1\u636e\u3002<\/p>\n<hr \/>\n<h3>\u5341\u56db\u3001\u8fd9\u7bc7\u8bba\u6587\u771f\u6b63\u7684\u521b\u65b0\u70b9\u662f\u4ec0\u4e48&#xff1f;<\/h3>\n<p>\u6211\u8ba4\u4e3a\u53ef\u4ee5\u603b\u7ed3\u6210\u56db\u70b9\u3002<\/p>\n<h4>14.1 \u7b2c\u4e00&#xff1a;\u91cd\u65b0\u5b9a\u4e49 CNN \u4e0e Attention \u4e4b\u95f4\u7684\u5dee\u522b<\/h4>\n<p>\u8bba\u6587\u6ca1\u6709\u7b80\u5355\u8bf4&#xff1a;<\/p>\n<p>Attention \u6bd4 CNN \u66f4\u5f3a&#xff0c;\u56e0\u4e3a Attention \u662f global \u7684<\/p>\n<p>\u800c\u662f\u8fdb\u4e00\u6b65\u4ece gated convolution \u7684\u5f62\u5f0f\u51fa\u53d1&#xff0c;\u6307\u51fa\u4e8c\u8005\u4e00\u4e2a\u975e\u5e38\u6838\u5fc3\u7684\u5dee\u5f02&#xff1a;<\/p>\n<p>Convolution&#xff1a;<\/p>\n<p>weight \u2190 relative position<\/p>\n<p>Self-Attention&#xff1a;<\/p>\n<p>weight \u2190 content similarity<\/p>\n<p>\u56e0\u6b64\u4f5c\u8005\u7684\u76ee\u6807\u5e76\u4e0d\u662f\u76f4\u63a5\u590d\u5236 Self-Attention&#xff0c;\u800c\u662f\u60f3\u529e\u6cd5\u8ba9&#xff1a;<\/p>\n<p>relative position<\/p>\n<p>\u80fd\u591f\u53cd\u6620&#xff1a;<\/p>\n<p>content similarity<\/p>\n<p>\u8fd9\u4e00\u7406\u8bba\u89c6\u89d2\u672c\u8eab\u5c31\u662f\u8bba\u6587\u6700\u91cd\u8981\u7684\u521b\u65b0\u4e4b\u4e00\u3002<\/p>\n<hr \/>\n<h4>14.2 \u7b2c\u4e8c&#xff1a;\u5229\u7528 Sorting \u628a Content Similarity \u53d8\u6210 Relative Position<\/h4>\n<p>\u8fd9\u662f\u8bba\u6587\u6700\u6838\u5fc3\u7684\u65b9\u6cd5\u521b\u65b0\u3002<\/p>\n<p>\u4f5c\u8005\u5229\u7528&#xff1a;<\/p>\n<p>Sort<\/p>\n<p>SortBy<\/p>\n<p>SortBack<\/p>\n<p>\u5efa\u7acb&#xff1a;<\/p>\n<p>Feature value similarity<\/p>\n<p>        \u2193<\/p>\n<p>Sorted positional proximity<\/p>\n<p>\u4e24\u4e2a\u672c\u6765 spatially distant \u7684 feature&#xff0c;\u53ea\u8981\u5185\u5bb9\u76f8\u4f3c&#xff0c;\u5c31\u53ef\u4ee5\u5728 sorted sequence \u4e2d\u6210\u4e3a\u90bb\u5c45\u3002\u968f\u540e\u666e\u901a 1D convolution \u5c31\u53ef\u4ee5\u6309\u7167 relative position \u5bf9\u5b83\u4eec\u8fdb\u884c\u805a\u5408\u3002\u56e0\u6b64\u4e0d\u9700\u8981\u663e\u5f0f\u8ba1\u7b97QK^T\u4e5f\u4e0d\u9700\u8981\u7ef4\u62a4<img decoding=\"async\" alt=\"N\\\\times N\" class=\"mathcode\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074949-6a7d771d80325.png\" \/>\u7684 pairwise relation matrix\u3002<\/p>\n<hr \/>\n<h4>14.3 \u7b2c\u4e09&#xff1a;Spatial &#043; Content \u53cc\u5206\u652f<\/h4>\n<p>\u4f5c\u8005\u6ca1\u6709\u56e0\u4e3a\u8981\u505a content modeling \u5c31\u629b\u5f03 CNN \u6700\u64c5\u957f\u7684 spatial inductive bias\u3002<\/p>\n<p>SpatialEgo \u540c\u65f6\u5305\u542b&#xff1a;<\/p>\n<p>7\u00d77 2D DWConv<\/p>\n<p>&#043;<\/p>\n<p>Sorting-based 1D DWConv<\/p>\n<p>\u524d\u8005\u8d1f\u8d23 local spatial relationship&#xff0c;\u540e\u8005\u8d1f\u8d23 content-driven long-range relationship\u3002<\/p>\n<p>\u56e0\u6b64 Ego \u5b9e\u9645\u4e0a\u8bd5\u56fe\u540c\u65f6\u4fdd\u7559&#xff1a;<\/p>\n<p>CNN \u7684 spatial modeling<br \/>\n&#043;<br \/>\nAttention \u7c7b\u4f3c\u7684 content-dependent interaction<\/p>\n<hr \/>\n<h4>14.4 \u7b2c\u56db&#xff1a;\u5c40\u90e8\u7a97\u53e3\u5b9e\u73b0\u9690\u5f0f Global Receptive Field<\/h4>\n<p>\u7531\u4e8e sorting \u4f1a\u91cd\u65b0\u6392\u5217\u6240\u6709 token&#xff0c;\u5373\u4f7f\u91c7\u752824\u8fd9\u6837\u7684 local window&#xff0c;\u4e5f\u53ef\u4ee5\u805a\u5408\u539f\u59cb\u56fe\u50cf\u4e2d\u7a7a\u95f4\u8ddd\u79bb\u5f88\u8fdc\u7684 feature\u3002<\/p>\n<p>\u6d88\u878d\u5b9e\u9a8c\u4e2d&#xff0c;\u5b83\u8fbe\u5230&#xff1a;<\/p>\n<p>84.0%<\/p>\n<p>\u4e0e2N-1 global convolution \u5b8c\u5168\u76f8\u540c\u3002\u56e0\u6b64\u4f5c\u8005\u6700\u7ec8\u5c06\u590d\u6742\u5ea6\u63a7\u5236\u5728<img decoding=\"async\" alt=\"O(N\\\\log N)\" class=\"mathcode\" src=\"2026-08-13tuiyrwwdofh.png\" \/>&#xff0c;\u800c\u4e0d\u662f\u663e\u5f0f pairwise modeling \u7684 quadratic complexity\u3002<\/p>\n<hr \/>\n<h3>\u5341\u4e94\u3001\u8bba\u6587\u4ecd\u7136\u6709\u54ea\u4e9b\u4e0d\u8db3&#xff1f;<\/h3>\n<p>\u4f5c\u8005\u5728 Conclusion \u4e2d\u4e3b\u52a8\u63d0\u51fa\u4e86\u51e0\u4e2a\u95ee\u9898\u3002<\/p>\n<h4>15.1 Sorting \u7684\u7406\u8bba\u89e3\u91ca\u8fd8\u4e0d\u591f\u5b8c\u6574<\/h4>\n<p>\u76ee\u524d\u8bba\u6587\u4e3b\u8981\u901a\u8fc7\u673a\u5236\u5206\u6790\u3001\u5b9e\u9a8c\u7ed3\u679c\u548c visualization \u6765\u8bf4\u660e&#xff1a;<\/p>\n<p>Sorting<br \/>\n\u2192<br \/>\nContent-driven aggregation<\/p>\n<p>\u662f\u6709\u6548\u7684\u3002<\/p>\n<p>\u4f46\u662f&#xff1a;<\/p>\n<p>\u4e3a\u4ec0\u4e48\u8fd9\u79cd\u57fa\u4e8e\u5355 channel feature value \u6392\u5e8f\u7684\u65b9\u6cd5\u80fd\u591f\u7a33\u5b9a\u8fd1\u4f3c\u6216\u8005\u66ff\u4ee3\u66f4\u52a0\u4e00\u822c\u7684 content similarity modeling&#xff1f;<\/p>\n<p>\u4f5c\u8005\u8ba4\u4e3a\u8fd8\u9700\u8981\u66f4\u5f3a\u7684 theoretical understanding\u3002<\/p>\n<p>\u8fd9\u4e00\u70b9\u786e\u5b9e\u4e5f\u662f\u8fd9\u7bc7\u8bba\u6587\u672a\u6765\u975e\u5e38\u503c\u5f97\u7814\u7a76\u7684\u95ee\u9898\u3002<\/p>\n<hr \/>\n<h4>15.2 Video \u548c Sequential Data \u4e2d\u53ef\u80fd\u51fa\u73b0\u4e25\u91cd\u95ee\u9898<\/h4>\n<p>\u5bf9\u4e8e static image&#xff1a;<\/p>\n<p>token \u987a\u5e8f\u88ab\u91cd\u65b0\u6392\u5217<\/p>\n<p>\u4e00\u822c\u4e0d\u4f1a\u9020\u6210\u65f6\u95f4\u56e0\u679c\u95ee\u9898\u3002<\/p>\n<p>\u4f46\u662f\u5982\u679c\u5e94\u7528\u5230&#xff1a;<\/p>\n<p>Video<\/p>\n<p>Time Series<\/p>\n<p>Sequential Data<\/p>\n<p>sorting \u5f88\u53ef\u80fd\u7834\u574f temporal order\u3002<\/p>\n<p>\u66f4\u4e25\u91cd\u7684\u662f&#xff0c;\u5728 causal task \u4e2d&#xff0c;\u5982\u679c\u672a\u6765 frame \u7684 feature \u7ecf\u8fc7 sorting \u88ab\u79fb\u52a8\u5230\u5f53\u524d frame \u9644\u8fd1&#xff0c;\u90a3\u4e48\u6a21\u578b\u53ef\u80fd\u95f4\u63a5\u4f7f\u7528&#xff1a;<\/p>\n<p>future information<\/p>\n<p>\u9020\u6210 information leakage\u3002\u56e0\u6b64\u4f5c\u8005\u8ba4\u4e3a\u672a\u6765\u9700\u8981\u7814\u7a76&#xff1a;<\/p>\n<p>Masked Sorting<\/p>\n<p>Time-aware Sorting<\/p>\n<p>\u7b49\u673a\u5236\u3002<\/p>\n<hr \/>\n<h4>15.3 \u53ef\u4ee5\u4e0e Deformable Convolution \u8fdb\u4e00\u6b65\u7ed3\u5408<\/h4>\n<p>\u4f5c\u8005\u6700\u540e\u8fd8\u63d0\u51fa&#xff0c;\u53ef\u4ee5\u7814\u7a76 sorted-sequence 1D convolution \u4e0e\u5176\u4ed6 convolution form \u7684\u7ed3\u5408&#xff0c;\u4f8b\u5982&#xff1a;<\/p>\n<p>Deformable Convolution<\/p>\n<p>\u56e0\u4e3a\u4e24\u8005\u5b9e\u9645\u4e0a\u90fd\u5728\u8bd5\u56fe\u7a81\u7834\u56fa\u5b9a local grid \u7684\u9650\u5236\u3002<\/p>\n<p>Deformable Conv&#xff1a;<\/p>\n<p>\u52a8\u6001\u6539\u53d8\u91c7\u6837\u4f4d\u7f6e<\/p>\n<p>Ego&#xff1a;<\/p>\n<p>\u52a8\u6001\u91cd\u6392 feature position<\/p>\n<p>\u4e8c\u8005\u7ed3\u5408\u4ee5\u540e&#xff0c;\u53ef\u80fd\u8fdb\u4e00\u6b65\u5e73\u8861&#xff1a;<\/p>\n<p>Local Structure Modeling<\/p>\n<p>\u548c<\/p>\n<p>Content-driven Aggregation<\/p>\n<hr \/>\n<h3>\u5341\u516d\u3001\u603b\u7ed3<\/h3>\n<p>\u8fd9\u7bc7\u8bba\u6587\u6700\u6709\u610f\u601d\u7684\u5730\u65b9\u5e76\u4e0d\u662f\u53c8\u63d0\u51fa\u4e86\u4e00\u79cd\u590d\u6742 Attention&#xff0c;\u800c\u662f\u91cd\u65b0\u601d\u8003\u4e86\u4e00\u4e2a\u975e\u5e38\u57fa\u7840\u7684\u95ee\u9898&#xff1a;<\/p>\n<p>\u5377\u79ef\u4e3a\u4ec0\u4e48\u53ea\u80fd\u6839\u636e\u7a7a\u95f4\u4f4d\u7f6e\u805a\u5408\u4fe1\u606f&#xff1f;<\/p>\n<p>\u4f5c\u8005\u53d1\u73b0&#xff0c;convolution parameter \u672c\u8eab\u786e\u5b9e\u53ea\u4f9d\u8d56 relative position&#xff0c;\u4f46\u662f\u5982\u679c\u80fd\u591f\u8ba9&#xff1a;<\/p>\n<p>Relative Position<\/p>\n<p>\u672c\u8eab\u643a\u5e26&#xff1a;<\/p>\n<p>Content Similarity<\/p>\n<p>\u90a3\u4e48\u4f20\u7edf convolution \u4e5f\u80fd\u591f\u95f4\u63a5\u5b8c\u6210 content-driven aggregation\u3002<\/p>\n<p>\u6240\u4ee5\u4ece\u6574\u7bc7\u8bba\u6587\u6765\u770b&#xff0c;\u5176\u771f\u6b63\u8d21\u732e\u53ef\u4ee5\u7528\u4e00\u53e5\u8bdd\u6982\u62ec&#xff1a;<\/p>\n<p>Ego \u6ca1\u6709\u8ba9 convolution \u5b66\u4f1a\u8ba1\u7b97 Self-Attention&#xff0c;\u800c\u662f\u901a\u8fc7\u6392\u5e8f\u6539\u53d8 token \u7684\u76f8\u5bf9\u4f4d\u7f6e&#xff0c;\u8ba9\u666e\u901a convolution \u81ea\u7136\u800c\u7136\u5730\u6309\u7167 content similarity \u8fdb\u884c\u4fe1\u606f\u805a\u5408\u3002<\/p>\n<p>\u6211\u8ba4\u4e3a\u8fd9\u4e2a\u601d\u8def\u6bd4\u5355\u7eaf\u201cCNN &#043; Attention\u201d\u66f4\u52a0\u503c\u5f97\u5173\u6ce8&#xff0c;\u56e0\u4e3a\u4f5c\u8005\u771f\u6b63\u4fee\u6539\u7684\u662f CNN \u4fe1\u606f\u805a\u5408\u673a\u5236\u4e2d\u201c\u4f4d\u7f6e\u201d\u548c\u201c\u5185\u5bb9\u201d\u4e4b\u95f4\u7684\u5173\u7cfb&#xff0c;\u800c\u4e0d\u662f\u7b80\u5355\u589e\u52a0\u4e00\u4e2a\u65b0\u7684\u6a21\u5757\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587\u9898\u76ee&#xff1a;Convolutional Neural Networks Driven by Content Similarity&#xff08;\u5185\u5bb9\u76f8\u4f3c\u5ea6\u9a71\u52a8\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc&#xff09;\u4f1a\u8bae&#xff1a;CVPR 2026\u6458\u8981&#xff1a;\u867d\u7136\u5377\u79ef\u795e\u7ecf\u7f51\u7edc(CNN)\u8fd1\u5e74\u6765\u4e0d\u65ad\u53d1\u5c55&#xff0c;\u4f46\u53d8\u5f62\u91d1\u521a\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u53d8\u5f97\u8d8a\u6765\u8d8a\u53d7\u6b22\u8fce\u3002\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d&#xff0c;\u6211\u4eec\u4ecb\u7ecd\u4e86\u4e00\u79cd\u57fa\u4e8e\u5185\u5bb9\u76f8\u4f3c\u5ea6\u7684CNN\u4fe1\u606f\u805a\u5408\u673a\u5236-\u4e00\u79cd\u7c7b\u4f3c\u4e8e\u81ea\u6211\u6ce8\u610f\u673a\u5236\u7684\u80fd\u529b\u3002\u4e3a\u6b64&amp;#xff<\/p>\n","protected":false},"author":2,"featured_media":94041,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[587,7620,2394,50],"topic":[],"class_list":["post-94061","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-587","tag-7620","tag-cnn","tag-50"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\uff08\u8bba\u6587\u901f\u8bfb\uff09Ego\uff1a\u5185\u5bb9\u76f8\u4f3c\u5ea6\u9a71\u52a8\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/94061.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\uff08\u8bba\u6587\u901f\u8bfb\uff09Ego\uff1a\u5185\u5bb9\u76f8\u4f3c\u5ea6\u9a71\u52a8\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u8bba\u6587\u9898\u76ee&#xff1a;Convolutional Neural Networks Driven by Content Similarity&#xff08;\u5185\u5bb9\u76f8\u4f3c\u5ea6\u9a71\u52a8\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc&#xff09;\u4f1a\u8bae&#xff1a;CVPR 2026\u6458\u8981&#xff1a;\u867d\u7136\u5377\u79ef\u795e\u7ecf\u7f51\u7edc(CNN)\u8fd1\u5e74\u6765\u4e0d\u65ad\u53d1\u5c55&#xff0c;\u4f46\u53d8\u5f62\u91d1\u521a\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u53d8\u5f97\u8d8a\u6765\u8d8a\u53d7\u6b22\u8fce\u3002\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d&#xff0c;\u6211\u4eec\u4ecb\u7ecd\u4e86\u4e00\u79cd\u57fa\u4e8e\u5185\u5bb9\u76f8\u4f3c\u5ea6\u7684CNN\u4fe1\u606f\u805a\u5408\u673a\u5236-\u4e00\u79cd\u7c7b\u4f3c\u4e8e\u81ea\u6211\u6ce8\u610f\u673a\u5236\u7684\u80fd\u529b\u3002\u4e3a\u6b64&amp;#xff\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/94061.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-08-13T07:49:50+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074948-6a7d771cb7721.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"9 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/94061.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/94061.html\",\"name\":\"\uff08\u8bba\u6587\u901f\u8bfb\uff09Ego\uff1a\u5185\u5bb9\u76f8\u4f3c\u5ea6\u9a71\u52a8\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-08-13T07:49:50+00:00\",\"dateModified\":\"2026-08-13T07:49:50+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/94061.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/94061.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/94061.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\uff08\u8bba\u6587\u901f\u8bfb\uff09Ego\uff1a\u5185\u5bb9\u76f8\u4f3c\u5ea6\u9a71\u52a8\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\uff08\u8bba\u6587\u901f\u8bfb\uff09Ego\uff1a\u5185\u5bb9\u76f8\u4f3c\u5ea6\u9a71\u52a8\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/94061.html","og_locale":"zh_CN","og_type":"article","og_title":"\uff08\u8bba\u6587\u901f\u8bfb\uff09Ego\uff1a\u5185\u5bb9\u76f8\u4f3c\u5ea6\u9a71\u52a8\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u8bba\u6587\u9898\u76ee&#xff1a;Convolutional Neural Networks Driven by Content Similarity&#xff08;\u5185\u5bb9\u76f8\u4f3c\u5ea6\u9a71\u52a8\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc&#xff09;\u4f1a\u8bae&#xff1a;CVPR 2026\u6458\u8981&#xff1a;\u867d\u7136\u5377\u79ef\u795e\u7ecf\u7f51\u7edc(CNN)\u8fd1\u5e74\u6765\u4e0d\u65ad\u53d1\u5c55&#xff0c;\u4f46\u53d8\u5f62\u91d1\u521a\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u53d8\u5f97\u8d8a\u6765\u8d8a\u53d7\u6b22\u8fce\u3002\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d&#xff0c;\u6211\u4eec\u4ecb\u7ecd\u4e86\u4e00\u79cd\u57fa\u4e8e\u5185\u5bb9\u76f8\u4f3c\u5ea6\u7684CNN\u4fe1\u606f\u805a\u5408\u673a\u5236-\u4e00\u79cd\u7c7b\u4f3c\u4e8e\u81ea\u6211\u6ce8\u610f\u673a\u5236\u7684\u80fd\u529b\u3002\u4e3a\u6b64&amp;#xff","og_url":"https:\/\/www.wsisp.com\/helps\/94061.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-08-13T07:49:50+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813074948-6a7d771cb7721.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"9 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/94061.html","url":"https:\/\/www.wsisp.com\/helps\/94061.html","name":"\uff08\u8bba\u6587\u901f\u8bfb\uff09Ego\uff1a\u5185\u5bb9\u76f8\u4f3c\u5ea6\u9a71\u52a8\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-08-13T07:49:50+00:00","dateModified":"2026-08-13T07:49:50+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/94061.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/94061.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/94061.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\uff08\u8bba\u6587\u901f\u8bfb\uff09Ego\uff1a\u5185\u5bb9\u76f8\u4f3c\u5ea6\u9a71\u52a8\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/94061","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=94061"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/94061\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/94041"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=94061"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=94061"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=94061"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=94061"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}