{"id":950,"date":"2024-03-17T19:36:50","date_gmt":"2024-03-17T11:36:50","guid":{"rendered":"http:\/\/www.tamanegi.xyz\/?p=950"},"modified":"2024-03-17T19:36:50","modified_gmt":"2024-03-17T11:36:50","slug":"top-k-visual-tokens-transformer","status":"publish","type":"post","link":"http:\/\/tamanegi.xyz\/?p=950","title":{"rendered":"Top-K Visual Tokens Transformer"},"content":{"rendered":"<h1>Abstract<\/h1>\n<p>\u53ef\u89c1\u5149\u6a21\u6001\u548c\u7ea2\u5916\u6a21\u6001\u7684\u4eba\u5458\u518d\u8bc6\u522b\uff08VI-ReID\uff09\u662f\u4e00\u4e2a\u975e\u5e38\u91cd\u8981\u4e14\u5177\u6709\u6311\u6218\u6027\u7684\u4efb\u52a1\u3002\u73b0\u6709\u7684\u5de5\u4f5c\u4e3b\u8981\u96c6\u4e2d\u5728\u4f7f\u7528\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08CNN\uff09\u6765\u51cf\u5c11\u6a21\u6001\u5dee\u5f02\u3002\u7136\u800c\uff0cCNN\u63d0\u53d6\u7684\u7279\u5f81\u53ef\u80fd\u5305\u542b\u65e0\u7528\u7684\u4e0e\u8eab\u4efd\u65e0\u5173\u7684\u4fe1\u606f\uff0c\u8fd9\u4e0d\u53ef\u907f\u514d\u5730\u964d\u4f4e\u4e86\u7279\u5f81\u7684\u533a\u5206\u5ea6\u3002\u4e3a\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u672c\u6587\u5f15\u5165\u4e86\u4e00\u4e2a\u540d\u4e3aTop-K Visual Tokens Transformer\uff08TVTR\uff09\u7684\u6846\u67b6\uff0c\u8be5\u6846\u67b6\u5229\u7528\u4e86\u4e00\u4e2atop-k\u89c6\u89c9\u4ee4\u724c\u9009\u62e9\u6a21\u5757\u6765\u51c6\u786e\u9009\u62e9\u524dk\u4e2a\u5177\u6709\u533a\u5206\u6027\u7684\u89c6\u89c9\u8865\u4e01\uff0c\u4ee5\u51cf\u5c11\u8eab\u4efd\u65e0\u5173\u4fe1\u606f\u7684\u5e72\u6270\u5e76\u5b66\u4e60\u5177\u6709\u533a\u5206\u6027\u7684\u7279\u5f81\u3002\u6b64\u5916\uff0c\u8fd8\u5f00\u53d1\u4e86\u4e00\u4e2a\u5168\u5c40-\u5c40\u90e8\u5706\u5f62\u635f\u5931\u51fd\u6570\uff0c\u7528\u4e8e\u4f18\u5316TVTR\u4ee5\u5b9e\u73b0\u8de8\u6a21\u6001\u6b63\u5411\u96c6\u4e2d\u548c\u8d1f\u5411\u5206\u79bb\u7279\u6027\u3002\u5728SYSU-MM01\u548cRegDB\u6570\u636e\u96c6\u4e0a\u7684\u5b9e\u9a8c\u7ed3\u679c\u8bc1\u660e\u4e86\u6211\u4eec\u65b9\u6cd5\u7684\u4f18\u8d8a\u6027\u3002<del>\u6e90\u4ee3\u7801\u5c06\u4f1a\u53d1\u5e03\u3002<\/del><\/p>\n<h1>Introduction<\/h1>\n<p>\u4eba\u5458\u518d\u8bc6\u522b\uff08ReID\uff09\u65e8\u5728\u901a\u8fc7\u4e0d\u540c\u6444\u50cf\u5934\u6355\u83b7\u7684\u56fe\u50cf\u5339\u914d\u540c\u4e00\u4eba\u5458[1]\u3002\u5927\u591a\u6570\u73b0\u6709\u7684\u4eba\u5458Re-ID\u65b9\u6cd5\u4e13\u6ce8\u4e8e\u8bc6\u522b\u7531\u53ef\u89c1\u6444\u50cf\u5934\u6355\u83b7\u7684\u76f8\u540c\u884c\u4eba\u56fe\u50cf\u3002\u7136\u800c\uff0c\u5728\u5149\u7ebf\u8f83\u5dee\u7684\u60c5\u51b5\u4e0b\uff0c\u53ef\u89c1\u6444\u50cf\u5934\u65e0\u6cd5\u6355\u83b7\u8db3\u591f\u7684\u4e00\u4e2a\u4eba\u7684\u4fe1\u606f\uff0c\u8fd9\u9650\u5236\u4e86\u5355\u6a21\u6001ReID\u5728\u5b9e\u9645\u76d1\u63a7\u4e2d\u7684\u9002\u7528\u6027\u3002\u4e3a\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u63d0\u51fa\u4e86\u4ea4\u53c9\u6a21\u6001\uff08\u53ef\u89c1-\u7ea2\u5916\uff09\u4eba\u5458\u518d\u8bc6\u522b\uff08VI-ReID\uff09\u4efb\u52a1\u3002VI-ReID\u4efb\u52a1\u65e8\u5728\u5339\u914d\u7ea2\u5916\uff08IR\uff09\u548c\u53ef\u89c1\u6444\u50cf\u5934\u6355\u83b7\u7684\u4eba\u5458\u56fe\u50cf[2]\u3002\u4e00\u4e9b\u57fa\u4e8ecnn\u7684\u65b9\u6cd5[1]\u5df2\u88ab\u63d0\u51fa\uff0c\u7528\u4e8e\u5bf9\u9f50\u4e0d\u540c\u6a21\u6001\u4e0b\u4e00\u4e2a\u4eba\u7684\u7279\u5f81\u5206\u5e03\u3002\u7136\u800c\uff0c\u884c\u4eba\u56fe\u50cf\u603b\u662f\u5305\u542b\u4e00\u4e9b\u65e0\u7528\u7684\u80cc\u666f\u4fe1\u606f\uff0c\u8fd9\u5bf9\u884c\u4eba\u518d\u8bc6\u522b\u7684\u6027\u80fd\u6709\u5bb3\u3002\u4e3a\u4e86\u66f4\u597d\u5730\u8fdb\u884c\u8de8\u6a21\u6001\u68c0\u7d22\uff0c\u6709\u5fc5\u8981\u5b9a\u4f4d\u51fa\u6700\u91cd\u8981\u7684\u4eba\u5458\u56fe\u50cf\u533a\u57df\uff0c\u5e76\u4f9d\u9760\u5b83\u4eec\u6765\u63d0\u53d6VI-ReID\u7684\u533a\u5206\u6027\u7279\u5f81\u3002<br \/>\n\u6700\u8fd1\uff0c\u89c6\u89c9transformer\uff08ViT\uff09[3, 4]\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\u4e0a\u53d6\u5f97\u4e86\u5de8\u5927\u6210\u529f\u3002ViT\u5305\u542b\u8bb8\u591a\u81ea\u6ce8\u610f\u529b\u5c42\u3002\u5728\u81ea\u6ce8\u610f\u529b\u673a\u5236\u4e2d\uff0c\u6bcf\u4e2a\u89c6\u89c9\u4ee4\u724c\u90fd\u4e0e\u7c7b\u522b\u4ee4\u724c\u76f8\u8fde\u3002\u6ce8\u610f\u529b\u94fe\u63a5\u7684\u5f3a\u5ea6\uff08\u6ce8\u610f\u529b\u5206\u6570\uff09\u53ef\u4ee5\u76f4\u89c2\u5730\u89e3\u91ca\u4e3a\u4ee4\u724c\u7684\u91cd\u8981\u6027\u5ea6\u91cf\u3002\u8fd9\u79cd\u7279\u6027\u53ef\u4ee5\u5e94\u7528\u4e8e\u5b9a\u4f4d\u6700\u91cd\u8981\u7684\u56fe\u50cf\u533a\u57df\uff0c\u4ece\u800c\u6709\u52a9\u4e8e\u5b66\u4e60\u533a\u5206\u6027\u7279\u5f81\u3002\u6709\u4e86\u8fd9\u4e9b\u533a\u5206\u6027\u7279\u5f81\uff0c\u5bf9\u9f50\u4e0d\u540c\u6a21\u6001\u7684\u7279\u5f81\u5c06\u66f4\u52a0\u9ad8\u6548\uff0c\u5982\u56fe1\u6240\u793a\u3002<br \/>\n![[Illustration of Top-K Visual Tokens Selection.png]][Top-K \u89c6\u89c9\u6807\u8bb0\u9009\u62e9\uff08TVTS\uff09\u7684\u56fe\u793a\uff0c\u7eff\u8272\u6846\u4ee3\u8868\u91cd\u8981\u533a\u57df\uff0c\u7ea2\u8272\u6846\u4ee3\u8868\u5305\u542b\u65e0\u7528\u7684\u4e0e\u8eab\u4efd\u65e0\u5173\u7684\u4fe1\u606f\uff08\u4f8b\u5982\u80cc\u666f\u4fe1\u606f\uff09\u7684\u533a\u57df]<\/p>\n<p>\u57fa\u4e8e\u4ee5\u4e0a\u52a8\u673a\uff0c\u672c\u6587\u4ecb\u7ecd\u4e86\u4e00\u79cd\u540d\u4e3aTop-K Visual Tokens Transformer\uff08TVTR\uff09\u7684\u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u5229\u7528\u4e86\u4e00\u4e2aTop-K Visual Tokens Selection module\uff08TVTS\uff09\u6765\u4ece\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u90e8\u9009\u62e9\u524dk\u4e2a\u91cd\u8981\u7684\u89c6\u89c9\u8865\u4e01(visual patches)\uff0c\u4ee5\u5b9a\u4f4d\u91cd\u8981\u7684\u56fe\u50cf\u533a\u57df\u5e76\u5b66\u4e60\u533a\u5206\u6027\u7279\u5f81\u3002\u4e3a\u4e86\u8fdb\u4e00\u6b65\u5904\u7406\u8de8\u6a21\u6001\u548c\u5185\u6a21\u6001\u53d8\u5316\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u5168\u5c40-\u5c40\u90e8\u5faa\u73af\uff08GLC\uff09\u635f\u5931\u6765\u4f18\u5316TVTR\uff0c\u5176\u4e2d\u5168\u5c40\u5faa\u73af\u635f\u5931\u4f18\u5316\u6837\u672c\u4e4b\u95f4\u7684\u8ddd\u79bb\u4e0e\u5206\u7c7b\u5668\uff0c\u5c40\u90e8\u5faa\u73af\u635f\u5931\u4f18\u5316\u4e00\u6279\u6837\u672c\u4e2d\u6837\u672c\u4e4b\u95f4\u7684\u8ddd\u79bb\u3002<br \/>\n\u672c\u6587\u4e3b\u8981\u8d21\u732e\u5982\u4e0b<\/p>\n<ul>\n<li>\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u4e2a\u540d\u4e3aTop-K Visual Tokens Transformer\uff08TVTR\uff09\u7684\u6846\u67b6\uff0c\u8be5\u6846\u67b6\u5229\u7528\u4e86Top-K Visual Tokens Selection\uff08TVTS\uff09\u6a21\u5757\uff0c\u51c6\u786e\u9009\u62e9\u524dk\u4e2a\u91cd\u8981\u7684\u89c6\u89c9\u8865\u4e01\uff0c\u4ee5\u51cf\u5c11\u8eab\u4efd\u65e0\u5173\u4fe1\u606f\u7684\u5e72\u6270\u5e76\u5b66\u4e60\u5177\u6709\u533a\u5206\u6027\u7684\u7279\u5f81\u3002<\/li>\n<li>\u6211\u4eec\u5f00\u53d1\u4e86\u4e00\u4e2a\u5168\u5c40-\u5c40\u90e8\u73af\u5f62\uff08GLC\uff09\u635f\u5931\u6765\u4f18\u5316TVTR\uff0c\u4ee5\u5b9e\u73b0\u8de8\u6a21\u6001\u6b63\u5411\u96c6\u4e2d\u548c\u8d1f\u5411\u5206\u79bb\u7684\u5c5e\u6027\u3002<\/li>\n<li>\u6211\u4eec\u5728SYSU-MM01\u548cRegDB\u6570\u636e\u96c6\u4e0a\u63d0\u4f9b\u4e86\u8be6\u7ec6\u7684\u5b9e\u9a8c\u7ed3\u679c\uff0c\u8bc1\u660e\u4e86\u6211\u4eec\u65b9\u6cd5\u7684\u4f18\u8d8a\u6027\u3002<\/li>\n<\/ul>\n<h1>Methodology<\/h1>\n<p>Top-K Visual Tokens Transformer\uff08TVTR\uff09\u7684\u6846\u67b6\u5982\u56fe2\u6240\u793a\u3002\u6211\u4eec\u9996\u5148\u5c06\u8f93\u5165\u7684\u53ef\u89c1\u6216\u7ea2\u5916\u4eba\u5458\u56fe\u50cf\u5207\u5272\u6210\u4e00\u7cfb\u5217\u91cd\u53e0\u7684\u6241\u5e73\u8865\u4e01\uff0c\u5e76\u4f7f\u7528\u7ebf\u6027\u6295\u5f71\u5c06\u8fd9\u4e9b\u8865\u4e01\u6620\u5c04\u5230D\u7ef4\u7279\u5f81\u7a7a\u95f4\u4e2d\u3002\u57fa\u672c\u7f16\u7801\u5668\u5305\u542b12\u5c42transformer\uff0c\u7531\u591a\u5934\u81ea\u6ce8\u610f\u529b\u548c\u591a\u5c42\u611f\u77e5\u5668\u5757\u7ec4\u6210\u3002\u6211\u4eec\u5728\u6700\u540e\u4e00\u5c42\u4e4b\u524d\u63d2\u5165\u4e86Top-K Visual Tokens Selection\uff08TVTS\uff09\u6a21\u5757\u3002\u6240\u9009\u7684\u89c6\u89c9\u4ee4\u724c\u548c\u7c7b\u522b\u4ee4\u724c\u88ab\u8f93\u5165\u5230\u6700\u540e\u4e00\u5c42\u8fdb\u884c\u6700\u7ec8\u5b66\u4e60\u3002<br \/>\n![[The framework of Top-K Visual Tokens Transformer..png]][TVTR \u7684\u6846\u67b6\u3002Top-K \u89c6\u89c9\u6807\u8bb0\u9009\u62e9\u53ef\u4ee5\u627e\u5230\u91cd\u8981\u7684 Visual Patches\u3002\u5168\u5c40-\u5c40\u90e8\u5faa\u574f\u635f\u5931\u7ed3\u5408\u4e86\u5168\u5c40\u548c\u5c40\u90e8\u5ea6\u91cf\u5b66\u4e60\uff0c\u5df2\u5b9e\u73b0\u8de8\u6a21\u6001\u7684\u6b63\u96c6\u4e2d\u548c\u8d1f\u5206\u79bb\u7279\u6027]<\/p>\n<h2>Top-K Visual Tokens Selection<\/h2>\n<p>\u5728transformer\u4e2d\uff0c\u6bcf\u4e2a\u89c6\u89c9\u4ee4\u724c\u90fd\u4e0e\u5206\u7c7b\u4ee4\u724c\u76f8\u8fde\u3002\u6ce8\u610f\u529b\u94fe\u63a5\u7684\u5f3a\u5ea6\u53ef\u4ee5\u76f4\u89c2\u5730\u89e3\u91ca\u4e3a\u89c6\u89c9\u4ee4\u724c\u7684\u91cd\u8981\u6027\u5ea6\u91cf\u3002Top-K Visual Tokens Selection\uff08TVTS\uff09\u6a21\u5757\u6574\u5408\u4e86\u6bcf\u4e2a\u81ea\u6ce8\u610f\u529b\u5c42\u7684\u6ce8\u610f\u529b\u5206\u6570\uff0c\u751f\u6210\u6700\u7ec8\u7684\u6ce8\u610f\u529b\u56fe\uff0c\u5f15\u5bfc\u6a21\u578b\u6709\u6548\u51c6\u786e\u5730\u9009\u62e9\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u90e8\u7684\u524dk\u4e2a\u5177\u6709\u533a\u5206\u6027\u7684\u56fe\u50cf\u8865\u4e01\u8fdb\u5165\u6700\u540e\u4e00\u5c42\uff0c\u4ee5\u83b7\u5f97\u6700\u7ec8\u7279\u5f81\u3002\u5047\u8bbe\u6a21\u578b\u5305\u542bM\u4e2a\u81ea\u6ce8\u610f\u529b\u5934\u90e8\uff0c\u5e76\u4e14\u8f93\u5165\u5230\u6700\u540e\u4e00\u5c42\u7684\u9690\u85cf\u7279\u5f81\u8868\u793a\u4e3a$H<em>{L-1}=[h^0<\/em>{L-1};h^1<em>{L-1},h^2<\/em>{L-1},...,h^N_{L-1}]$\uff0c\u5176\u4e2dN\u662f\u56fe\u50cf\u8865\u4e01\u7684\u6570\u91cf\uff0cL\u662f\u5c42\u6570\u3002\u5219\u524d\u9762\u5c42\u7684\u6ce8\u610f\u529b\u6743\u91cd\u53ef\u4ee5\u8868\u793a\u4e3a\uff1a<br \/>\n$$<br \/>\nA_l=[a^0_l,a^1_l.a^2_l,...,a^M_l]\\qquad l\\in1,2,...,L-1<br \/>\n$$<br \/>\n\u5176\u4e2d$a^i_l$\u662f\u7b2c$l$\u5c42\u7684\u7b2c$i$\u4e2a\u6ce8\u610f\u529b\u5934\u7684\u6743\u91cd\uff0c\u53ef\u4ee5\u5199\u6210\uff1a<br \/>\n$$<br \/>\na^i_l=[a^{i_0}_l;a^{i_1}_l,a^{i_2}_l,...,a^{i_N}_l] \\qquad i\\in0,1,...,M<br \/>\n$$<br \/>\n\u5176\u4e2d$i_N$\u662f\u4e00\u5f20\u56fe\u7684\u7b2c N \u4e2a patch\u3002$i<em>0$\u662f\u7c7b\u522b token\uff0c\u6ce8\u610f\u529b\u6743\u91cd\u53ef\u4ee5\u88ab\u8ba1\u7b97\u4e3a:<br \/>\n$$<br \/>\n\\begin{aligned}<br \/>\na(Q,K)=softmax(\\frac{QK^T}{\\sqrt d})\\<br \/>\n\\end{aligned}<br \/>\n$$<br \/>\n\u5176\u4e2d$\\frac{1}{\\sqrt d}$\u662f\u4e00\u4e2a\u6bd4\u4f8b\u56e0\u5b50\u3002<br \/>\n\u4e3a\u4e86\u4f7f\u5d4c\u5165\u53ef\u8bc6\u522b\uff0c\u6211\u4eec\u6574\u5408\u4e86\u6240\u6709\u524d\u51e0\u5c42\u7684\u6ce8\u610f\u529b\u6743\u91cd\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6700\u7ec8\u7684\u6ce8\u610f\u529b\u5206\u6570\u53ef\u4ee5\u901a\u8fc7\u77e9\u9635\u4e58\u6cd5\u9012\u5f52\u8ba1\u7b97\u5982\u4e0b\uff1a<br \/>\n$$<br \/>\nA<\/em>{final}=\\prod\\limits^{L-1}_{l=1}A<em>l<br \/>\n$$<br \/>\n$A<\/em>{final}$\u7684\u5143\u7d20\u8868\u793a\u4e24\u4e2a\u5bf9\u5e94patch\u4e4b\u95f4\u7684\u5173\u8054\u5f3a\u5ea6\u3002$A<em>{final}$\u53ef\u4ee5\u4f5c\u4e3a\u9009\u62e9\u5177\u6709\u533a\u5206\u6027\u533a\u57df\u7684\u6307\u6807\u3002\u7136\u540e\uff0c\u6211\u4eec\u9009\u62e9\u4e0e\u7c7b\u4ee4\u724c\u76f8\u5173\u8054\u7684\u524dk\u4e2a\u6700\u5927\u503c\u7684\u4f4d\u7f6e\u3002\u8fd9\u4e9b\u4f4d\u7f6e\u88ab\u7528\u4f5cTVTR\u4e2d\u63d0\u53d6$H<\/em>{L-1}$\u4e2d\u76f8\u5e94\u4ee4\u724c\u7684\u7d22\u5f15\u3002\u6240\u9009\u4ee4\u724c\u4e0e\u7c7b\u4ee4\u724c\u4e00\u8d77\u8f93\u5165\u5230\u6700\u540e\u4e00\u5c42\u3002\u6700\u540e\uff0c\u6211\u4eec\u5c06\u7c7b\u4ee4\u724c\u548c\u6700\u540e\u4e00\u5c42\u7684\u89c6\u89c9\u4ee4\u724c\u7684\u5168\u5c40\u5e73\u5747\u6c60\u8fde\u63a5\u5728\u4e00\u8d77\uff0c\u4f5c\u4e3a\u4e00\u4e2a\u4eba\u7684\u8868\u793a\u3002<\/p>\n<h2>Global-Local Circle Loss<\/h2>\n<p>Circle loss \u5728\u5404\u79cd\u6df1\u5ea6\u7279\u5f81\u5b66\u4e60\u4efb\u52a1\u4e0a\u5b9e\u73b0\u4e86\u4f18\u8d8a\u7684\u6027\u80fd\u3002\u4e3a\u4e86\u540c\u65f6\u5904\u7406\u8de8\u6a21\u6001\u548c\u5185\u6a21\u6001\u53d8\u5316\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u4e2a Global-Local Circle (GLC) loss\uff0c\u5305\u62ec\u4e24\u79cd\u7c7b\u578b\u7684\u5706\u5f62\u635f\u5931\u7528\u4e8e\u8de8\u6a21\u6001\u5b66\u4e60\uff0c\u5373\u5168\u5c40\u5706\u5f62\u635f\u5931\u548c\u5c40\u90e8\u5706\u5f62\u635f\u5931\u3002\u5168\u5c40\u5706\u5f62\u635f\u5931\u7528\u4e8e\u4f18\u5316\u6837\u672c\u4e0e\u5206\u7c7b\u5668\u4e4b\u95f4\u7684\u8ddd\u79bb\uff0c\u53ef\u4ee5\u89c6\u4e3a\u5168\u5c40\u951a\u70b9\u3002\u5c40\u90e8\u5706\u5f62\u635f\u5931\u7528\u4e8e\u4f18\u5316\u4e00\u4e2a\u6279\u6b21\u4e2d\u6837\u672c\u4e4b\u95f4\u7684\u8ddd\u79bb\u3002\u7ed3\u5408\u5168\u5c40\u548c\u5c40\u90e8\u5ea6\u91cf\u5b66\u4e60\uff0c\u53ef\u4ee5\u5b66\u4e60\u5230\u66f4\u597d\u7684\u7279\u5f81\uff0c\u5b9e\u73b0\u8de8\u6a21\u6001\u6b63\u96c6\u4e2d\u548c\u8d1f\u5206\u79bb\u7684\u7279\u6027\u3002\u5168\u5c40\u5706\u5f62\u635f\u5931\u8868\u793a\u4e3a\uff1a<br \/>\n$$L_g=-\\log\\frac{\\exp(\\gamma\\alpha^i_p(s_p-\\Delta_p))}{\\exp(\\gamma\\alpha^i_p(s_p-\\Delta<em>p))+\\sum^{N-1}<\/em>{j=1}\\exp(\\gamma\\alpha^j_n(s^j_n-\\Delta_n))}$$<br \/>\n\u5176\u4e2d$s^j_n$\u662f\u7c7b\u95f4\u76f8\u4f3c\u5ea6\u5f97\u5206\uff0c\u53ef\u4ee5\u7528$s^j_n=\\omega^T_jx\/(|\\omega_j||x|)$\u8ba1\u7b97\u5f97\u5230,$s_p$\u662f\u7c7b\u5185\u76f8\u4f3c\u5ea6\u5206\u6570\uff0c\u53ef\u4ee5\u7528$s_p=\\omega^T_yx\/(|\\omega_y||x|)$\u8ba1\u7b97\u5f97\u5230\u3002$w_j$ \u548c $w_y$ \u5206\u522b\u662f\u5206\u7c7b\u5668\u6743\u91cd\u4e2d\u7684\u975e\u76ee\u6807\u6743\u91cd\u5411\u91cf\u548c\u76ee\u6807\u6743\u91cd\u5411\u91cf\u3002$L_g$ \u7c7b\u4f3c\u4e8e Softmax \u6216 AMSoftmax \u635f\u5931\uff0c\u4f46\u5e26\u6709\u5706\u5f62\u52a0\u6743\u56e0\u5b50 ($a^i_p$ \u548c $a^j<em>n$)\u3002\u5bf9\u4e8e\u8de8\u6a21\u6001\u5b66\u4e60\uff0c\u5c40\u90e8\u5706\u5f62\u635f\u5931\u7ed3\u5408\u4e86\u57fa\u672c\u5706\u5f62\u635f\u5931\u7684\u4e09\u4e2a\u53d8\u4f53\uff0c\u5305\u62ec\u7ea2\u5916\u5706\u5f62\u635f\u5931\u3001\u53ef\u89c1\u5149\u5706\u5f62\u635f\u5931\u548c\u7ea2\u5916-\u53ef\u89c1\u5149\u5706\u5f62\u635f\u5931\u3002\u5c40\u90e8\u5706\u5f62\u635f\u5931\u53ef\u4ee5\u8868\u793a\u4e3a:<br \/>\n$$\\mathcal{L}<\/em>{1}=\\mathcal{L}<em>{ii}+\\mathcal{L}<\/em>{vv}+\\mathcal{L}<em>{iv}$$<br \/>\n$\\mathcal{L}<\/em>{ii},\\mathcal{L}<em>{vv},\\mathcal{L}<\/em>{iv}$\u5206\u522b\u662f\u7ea2\u5916\u5706\u635f\u5931\uff0c\u53ef\u89c1\u5149\u5706\u635f\u5931\u548c\u7ea2\u5916-\u53ef\u89c1\u5149\u5706\u635f\u5931\u3002\u7ea2\u5916\u5706\u5f62\u635f\u5931\u548c\u53ef\u89c1\u5149\u5706\u5f62\u635f\u5931\u7c7b\u4f3c\u4e8e\u57fa\u672c\u7684\u5706\u5f62\u635f\u5931\uff0c\u4f18\u5316\u5355\u6a21\u6001\u6837\u672c\u4e4b\u95f4\u7684\u8ddd\u79bb\uff0c\u5176\u8868\u8fbe\u5f0f\u4e3a\uff1a<br \/>\n$$<br \/>\n\\begin{aligned}<br \/>\n\\mathcal{L}<em>{vv}=log[1+\\sum\\limits^{L}<\/em>{i=1}\\exp(\\gamma\\alpha^j<em>n(s^j<\/em>{n(vv)}-\\Delta<em>n))\\sum\\limits^{K}<\/em>{i=1}\\exp(-\\gamma\\alpha^i<em>P(s^i<\/em>{p(vv)}-\\Delta_p))]\\<\/p>\n<p>\\mathcal{L}<em>{ii}=log[1+\\sum\\limits^{L}<\/em>{i=1}\\exp(\\gamma\\alpha^j<em>n(s^j<\/em>{n(ii)}-\\Delta<em>n))\\sum\\limits^{K}<\/em>{i=1}\\exp(-\\gamma\\alpha^i<em>P(s^i<\/em>{p(ii)}-\\Delta_p))]<\/p>\n<p>\\end{aligned}<br \/>\n$$<br \/>\n\u5176\u4e2d$s<em>{p(vv)},s<\/em>{n(vv)}$\u662f\u53ef\u89c1\u5149\u6a21\u6001\u4e2d\u6b63\u8d1f\u6837\u672c\u7684\u4f59\u5f26\u76f8\u4f3c\u5ea6\u3002\u540c\u6837\u7684\uff0c$s<em>{p(ii)},s<\/em>{n(ii)}$\u662f\u7ea2\u5916\u5149\u6a21\u6001\u4e2d\u6b63\u8d1f\u6837\u672c\u7684\u4f59\u5f26\u76f8\u4f3c\u5ea6\u3002<br \/>\n\u7ea2\u5916-\u53ef\u89c1\u5149\u5706\u5f62\u635f\u5931\u8868\u793a\u4e3a\uff1a<br \/>\n$$<br \/>\n\\begin{aligned}<br \/>\n\\mathcal{L}<em>{iv}=log[1+\\sum\\limits^{L}<\/em>{i=1}\\exp(\\gamma\\alpha^j<em>n(s^j<\/em>{n(iv)}-\\Delta<em>n))\\sum\\limits^{K}<\/em>{i=1}\\exp(-\\gamma\\alpha^i<em>P(s^i<\/em>{p(iv)}-\\Delta<em>p))]<br \/>\n\\end{aligned}<br \/>\n$$<br \/>\n\u5176\u4e2d\uff0c$s<\/em>{n(iv)},s_{p(iv)}$\u662f\u8de8\u6a21\u6001\u4e0b\u6b63\u8d1f\u6837\u672c\u7684\u4f59\u5f26\u76f8\u4f3c\u5ea6\u3002$\\Delta_p=1-m,\\Delta_n=m$\u63a7\u5236\u4f59\u5f26\u677e\u5f1b\u5ea6\u3002$\\alpha^i_p,\\alpha^j<em>n$\u7684\u5b9a\u4e49\u53ef\u4ee5\u4ece[Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei, \u201cCircle loss: A unified perspective of pair similarity optimization,\u201d in Proceedings of the IEEE\/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 6398\u20136407.]\u4e2d\u627e\u5230\uff08\uff1f\uff09\u3002<br \/>\nTVTR \u7684\u603b\u4f53\u635f\u5931\u53ef\u4ee5\u8868\u793a\u4e3a\uff1a<br \/>\n$$\\mathcal{L}<\/em>{overall}=\\mathcal{L}_l+\\mathcal{L}_g$$<\/p>\n<h3>\u8ba8\u8bba<\/h3>\n<p>\u4ee5\u5f80\u7684ReID\u5de5\u4f5c\u901a\u5e38\u4f7f\u7528\u4e09\u5143\u7ec4\u635f\u5931\u548c\u4ea4\u53c9\u71b5\u635f\u5931\u6765\u4f18\u5316\u6a21\u578b\u3002GLC\u635f\u5931\u63d0\u4f9b\u4e86\u4e24\u79cd\u5ea6\u91cf\u5b66\u4e60\u7684\u65b9\u5f0f\uff0c\u7c7b\u4f3c\u4e8e\u4e09\u5143\u7ec4\u635f\u5931\u548c\u4ea4\u53c9\u71b5\u635f\u5931\u7684\u7ed3\u5408\uff0c\u5305\u62ec\u4f18\u5316\u6837\u672c\u4e0e\u5206\u7c7b\u5668\u4e4b\u95f4\u7684\u8ddd\u79bb\uff08\u5168\u5c40\u5706\u5f62\u635f\u5931\uff09\u548c\u4e00\u6279\u6837\u672c\u4e4b\u95f4\u7684\u8ddd\u79bb\uff08\u5c40\u90e8\u5706\u5f62\u635f\u5931\uff09\u3002\u901a\u8fc7\u7ed3\u5408\u5168\u5c40\u548c\u5c40\u90e8\u635f\u5931\uff0c\u6a21\u578b\u53ef\u4ee5\u5b66\u4e60\u5230\u66f4\u597d\u7684\u5d4c\u5165\uff0c\u7528\u4e8e\u8de8\u6a21\u6001\u68c0\u7d22\uff0c\u51cf\u5c0f\u7c7b\u5185\u8ddd\u79bb\uff0c\u6269\u5927\u7c7b\u95f4\u8ddd\u79bb\u3002<\/p>\n<h1>Experiments<\/h1>\n<h2>Datasets<\/h2>\n<h2>Implementation Details<\/h2>\n<p>\u6211\u4eec\u63d0\u51fa\u7684\u65b9\u6cd5\u662f\u4f7f\u7528PyTorch\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u5b9e\u73b0\u7684\u3002\u6211\u4eec\u91c7\u7528\u4e86\u5728ImageNet\u6570\u636e\u96c6\u4e0a\u9884\u8bad\u7ec3\u7684ViT-Base\u6a21\u578b\u4f5c\u4e3atransformer\u7684\u4e3b\u5e72\u3002\u53ef\u89c1\u5149\u548c\u7ea2\u5916\u56fe\u50cf\u5728\u8fdb\u5165\u7f51\u7edc\u4e4b\u524d\u88ab\u7f29\u653e\u4e3a3\u00d7288\u00d7144\uff08C \u00d7 H \u00d7 W\uff09\u3002\u5728top-k\u89c6\u89c9\u6807\u8bb0\u9009\u62e9\u4e2d\uff0c\u6bcf\u4e2a\u5934\u90e8\u4e2d\u7684\u524d16\u4e2a\u91cd\u8981\u89c6\u89c9\u6807\u8bb0\u88ab\u9009\u62e9\u5230\u6700\u540e\u4e00\u5c42\u3002\u91c7\u7528\u4e86\u5e26\u6709\u6743\u91cd\u8870\u51cf1e-4\u7684SGD\u4f18\u5316\u5668\u3002\u5168\u5c40-\u5c40\u90e8\u5708\u635f\u5931\u6709\u4e09\u4e2a\u8d85\u53c2\u6570\uff0c\u5373\u03b3\uff0c\u2206p\u548c\u2206n\u3002\u5bf9\u4e8e\u5168\u5c40\u5708\u635f\u5931\uff0c\u03b3 = 30\u3002\u2206p\u548c\u2206n\u90fd\u8bbe\u7f6e\u4e3a0.5\u3002\u5bf9\u4e8e\u5c40\u90e8\u5708\u635f\u5931\uff0c\u03b3 = 20\u3002\u2206p\u8bbe\u7f6e\u4e3a0.75\uff0c\u2206n\u8bbe\u7f6e\u4e3a0.25\u3002<\/p>\n<h2>Comparison Results<\/h2>\n<p>\u6211\u4eec\u5728SYSU-MM01\u548cRegDB\u6570\u636e\u96c6\u4e0a\u5c06\u6211\u4eec\u7684\u65b9\u6cd5\u4e0e\u6700\u5148\u8fdb\u7684\u7b97\u6cd5\u8fdb\u884c\u4e86\u6bd4\u8f83\uff0c\u7ed3\u679c\u89c1\u88681\u3002\u6211\u4eec\u63d0\u51fa\u7684TVTR\u6846\u67b6\u5728\u6027\u80fd\u4e0a\u660e\u663e\u4f18\u4e8e\u73b0\u6709\u7684\u57fa\u4e8eCNN\u7684\u65b9\u6cd5\uff0c\u8fd9\u8bc1\u660e\u4e86\u6211\u4eec\u65b9\u6cd5\u7684\u6709\u6548\u6027\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5728SYSU-MM01\u6570\u636e\u96c6\u4e0a\uff0c\u6211\u4eec\u7684TVTR\u5728mAP\u65b9\u9762\u8fbe\u5230\u4e8665.30%\uff0c\u5728Rank-1\u65b9\u9762\u8fbe\u5230\u4e8664.15%\uff08All-Search\u6a21\u5f0f\uff09\uff0c\u5206\u522b\u6bd4DG-VAE\u63d0\u9ad8\u4e865.81%\u548c5.69%\u3002\u4e0e\u53e6\u4e00\u79cd\u57fa\u4e8etransformer\u7684\u65b9\u6cd5CMTR\u76f8\u6bd4\uff0c\u6211\u4eec\u7684\u65b9\u6cd5\u5728mAP\u51c6\u786e\u5ea6\u548cRank-1\u4e0a\u5206\u522b\u63d0\u9ad8\u4e862.72%\u548c2.82%\u3002\u5728RegDB\u6570\u636e\u96c6\u4e0a\uff0c\u6211\u4eec\u5728\u53ef\u89c1\u5149\u5230\u70ed\u7ea2\u5916\u6d4b\u8bd5\u6a21\u5f0f\u4e0b\u5b9e\u73b0\u4e8679.5%\u7684mAP\u548c84.1%\u7684Rank-1\u3002\u4e0eCMTR\u76f8\u6bd4\uff0c\u6211\u4eec\u7684\u65b9\u6cd5\u5728mAP\u51c6\u786e\u5ea6\u548cRank-1\u4e0a\u5206\u522b\u63d0\u9ad8\u4e863.5%\u548c5.1%\u3002\u8fd9\u4e9b\u7ed3\u679c\u8bc1\u660e\u4e86\u6211\u4eec\u63d0\u51fa\u7684TVTR\u6846\u67b6\u7684\u6301\u7eed\u6539\u8fdb\u3002<\/p>\n<h2>Ablation Studies<\/h2>\n<p>\u6211\u4eec\u901a\u8fc7\u5728SYSU-MM01\u4e0a\u8fdb\u884c\u6d88\u878d\u7814\u7a76\u6765\u8bc4\u4f30\u6211\u4eec\u63d0\u51fa\u7684TVTR\u6846\u67b6\u4e2d\u7684\u6bcf\u4e2a\u7ec4\u4ef6\u3002\u7ed3\u679c\u5982\u88682\u6240\u793a\u3002<\/p>\n<h3>Baseline<\/h3>\n<p>\u57fa\u7ebf\u8868\u793a\u76f4\u63a5\u901a\u8fc7\u901a\u7528 Lid \u548c Lwrt [1] \u5bf9 ViT \u6a21\u578b\u8fdb\u884c\u8bad\u7ec3\uff0c\u7528\u4e8e\u8de8\u6a21\u6001 ReID\u3002\u53ef\u4ee5\u89c2\u5bdf\u5230\uff0c\u57fa\u7ebf\u5728 mAP \u65b9\u9762\u8fbe\u5230\u4e86 51.26%\uff0c\u5728 Rank-1 \u548c mAP \u65b9\u9762\u5206\u522b\u6bd4 AGW \u63d0\u9ad8\u4e86 4.16% \u548c 3.61%\u3002<\/p>\n<h3>Effectiveness of Top-K Visual Tokens Selection<\/h3>\n<p>\u5982\u88682\u6240\u793a\uff0c\u901a\u8fc7\u5e94\u7528 TVTS \u6a21\u5757\u9009\u62e9\u524d k \u4e2a\u5177\u6709\u533a\u5206\u6027\u7684\u89c6\u89c9 token\uff0c\u6a21\u578b\u7684\u6027\u80fd\u4ece 51.26% \u63d0\u9ad8\u5230\u4e86 55.62% \u7684 mAP\u3002TVTS \u6a21\u5757\u76f4\u63a5\u4e22\u5f03\u4e86\u4e00\u4e9b\u65e0\u7528\u7684 token\uff0c\u5e76\u8feb\u4f7f\u7f51\u7edc\u4ece\u91cd\u8981\u7684\u56fe\u50cf\u90e8\u5206\u8fdb\u884c\u5b66\u4e60\u3002<\/p>\n<h3>Effectiveness of Global-Local Circle Loss<\/h3>\n<p>GLC \u635f\u5931\u6709\u52a9\u4e8e\u6a21\u578b\u4ece\u5c40\u90e8\u548c\u5168\u5c40\u4f18\u5316\u7684\u89d2\u5ea6\u964d\u4f4e\u7279\u5f81\u5d4c\u5165\u7ea7\u522b\u7684\u7c7b\u5185\u5dee\u5f02\uff0c\u5e76\u6269\u5927\u7c7b\u95f4\u5dee\u5f02\u3002\u5728 SYSU-MM01 \u6570\u636e\u96c6\u4e0a\uff0c\u6027\u80fd\u663e\u8457\u63d0\u9ad8\u4e86 8.97% \u7684 Rank-1 \u548c 9.20% \u7684 mAP\u3002<\/p>\n<h3>Evaluation of Hyperparameter K<\/h3>\n<p>\u6211\u4eec\u5728 SYSU-MM01 \u6570\u636e\u96c6\u4e0a\u8bc4\u4f30\u4e86 TVTS \u6a21\u5757\u4e2d\u4e0d\u540c K \u503c\u7684\u5f71\u54cd\uff0c\u91c7\u7528\u5168\u641c\u7d22\u6a21\u5f0f\u3002\u6211\u4eec\u5c06 K \u4ece 4 \u53d8\u5316\u5230 24\u3002\u5982\u56fe 3 \u6240\u793a\uff0c\u5f53 K = 16 \u65f6\uff0c\u6a21\u578b\u8fbe\u5230\u4e86\u6700\u9ad8\u7684 Rank-1 \u7cbe\u5ea6\u3002<\/p>\n<h3>Visualization Analysis<\/h3>\n<p>\u6211\u4eec\u5728\u56fe 4 \u4e2d\u7ed8\u5236\u4e86\u6765\u81ea SYSU-MM01 \u6570\u636e\u96c6\u4e2d\u968f\u673a\u9009\u53d6\u7684 10 \u4e2a\u8eab\u4efd\u7684 t-SNE \u56fe\u3002\u6211\u4eec\u89c2\u5bdf\u5230\uff0c\u5728\u521d\u59cb\u9636\u6bb5\uff0c\u4e24\u79cd\u6a21\u6001\u7684\u7279\u5f81\u662f\u5206\u5f00\u7684\u3002\u7ecf\u8fc7\u8bad\u7ec3\u540e\uff0c\u6765\u81ea\u4e24\u79cd\u6a21\u6001\u7684\u6bcf\u4e2a\u8eab\u4efd\u7684\u7279\u5f81\u5728\u5b66\u4e60\u5230\u7684\u5d4c\u5165\u7a7a\u95f4\u4e2d\u805a\u96c6\u5728\u4e00\u8d77\u3002<\/p>\n<h1>Conclusion<\/h1>\n<p>\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd Top-K Visual Token Transformer (TVTR) \u6846\u67b6\uff0c\u5229\u7528 top-k \u89c6\u89c9\u4ee4\u724c\u9009\u62e9\u6a21\u5757\u51c6\u786e\u9009\u62e9\u524d k \u4e2a\u5177\u6709\u533a\u5206\u6027\u7684\u89c6\u89c9\u8865\u4e01\uff0c\u4ee5\u51cf\u5c11\u4e0e\u8eab\u4efd\u65e0\u5173\u4fe1\u606f\u7684\u5e72\u6270\u3002\u6b64\u5916\uff0c\u6211\u4eec\u5f00\u53d1\u4e86\u4e00\u79cd\u5168\u5c40-\u5c40\u90e8\u5706\u5f62\u635f\u5931\u6765\u4f18\u5316\u6837\u672c\u4e0e\u5206\u7c7b\u5668\u4e4b\u95f4\u7684\u8ddd\u79bb\uff08\u5168\u5c40\u5706\u5f62\u635f\u5931\uff09\u4ee5\u53ca\u4e00\u6279\u6837\u672c\u4e4b\u95f4\u7684\u8ddd\u79bb\uff08\u5c40\u90e8\u5706\u5f62\u635f\u5931\uff09\u3002\u5b9e\u9a8c\u8bc1\u660e\u4e86\u6211\u4eec\u65b9\u6cd5\u7684\u6709\u6548\u6027\uff0c\u8868\u660e\u57fa\u4e8e Transformer \u7684\u6a21\u578b\u5728 VI-ReID \u4efb\u52a1\u4e0a\u5177\u6709\u5de8\u5927\u6f5c\u529b\u3002\u5728\u672a\u6765\u7684\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u5c06\u63d0\u4f9b\u89c6\u89c9\u793a\u4f8b\uff0c\u5c55\u793a\u54ea\u4e9b\u56fe\u50cf\u8865\u4e01\u5bf9\u4e8e\u8bc6\u522b\u4e00\u4e2a\u4eba\u662f\u91cd\u8981\u7684\uff0c\u5e76\u7ed9\u51fa\u5c06 TVTS \u96c6\u6210\u5230 ViT \u7684\u4e2d\u95f4\u5c42\u7684\u7ed3\u679c\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Abstract \u53ef\u89c1\u5149\u6a21\u6001\u548c\u7ea2\u5916\u6a21\u6001\u7684\u4eba\u5458\u518d\u8bc6\u522b\uff08VI-ReID\uff09\u662f\u4e00\u4e2a\u975e\u5e38\u91cd\u8981\u4e14\u5177\u6709\u6311\u6218\u6027\u7684\u4efb\u52a1\u3002\u73b0\u6709\u7684\u5de5 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[105,11],"tags":[],"_links":{"self":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts\/950"}],"collection":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=950"}],"version-history":[{"count":1,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts\/950\/revisions"}],"predecessor-version":[{"id":951,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts\/950\/revisions\/951"}],"wp:attachment":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=950"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=950"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=950"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}