{"id":21595,"date":"2024-08-01T09:49:44","date_gmt":"2024-08-01T01:49:44","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=21595"},"modified":"2024-08-01T09:49:44","modified_gmt":"2024-08-01T01:49:44","slug":"%e7%9b%ae%e6%a0%87%e6%a3%80%e6%b5%8b%e6%96%b0%e6%96%b9%e5%bc%8f-class-agnostic%e6%a3%80%e6%b5%8b%e5%99%a8%e7%94%a8%e4%ba%8e%e7%9b%ae%e6%a0%87%e6%a3%80%e6%b5%8b","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2024\/08\/01\/21595\/","title":{"rendered":"\u76ee\u6807\u68c0\u6d4b\u65b0\u65b9\u5f0f | class-agnostic\u68c0\u6d4b\u5668\u7528\u4e8e\u76ee\u6807\u68c0\u6d4b"},"content":{"rendered":"
\"\"<\/div>\n
\u76ee\u6807\u68c0\u6d4b\u6a21\u578b\u5728\u5b9a\u4f4d\u548c\u5206\u7c7b\u8bad\u7ec3\u671f\u95f4\u663e\u793a\u76ee\u6807\u65f6\u8868\u73b0\u826f\u597d\uff0c\u7136\u800c\uff0c\u7531\u4e8e\u521b\u5efa\u548c\u6ce8\u91ca\u68c0\u6d4b\u6570\u636e\u96c6\u7684\u96be\u5ea6\u548c\u6210\u672c\uff0c\u8bad\u7ec3\u8fc7\u7684\u6a21\u578b\u68c0\u6d4b\u5230\u6570\u91cf\u6709\u9650\u7684\u76ee\u6807\u7c7b\u578b\uff0c\u672a\u77e5\u76ee\u6807\u88ab\u89c6\u4e3a\u80cc\u666f\u5185\u5bb9\u3002\u8fd9\u963b\u788d\u4e86\u4f20\u7edf\u68c0\u6d4b\u5668\u5728\u73b0\u5b9e\u5e94\u7528\u4e2d\u7684\u91c7\u7528\uff0c\u5982\u5927\u89c4\u6a21\u7269\u4f53\u5339\u914d\u3001visual grounding\u3001\u89c6\u89c9\u5173\u7cfb\u9884\u6d4b\u3001\u969c\u788d\u68c0\u6d4b\uff08\u786e\u5b9a\u7269\u4f53\u7684\u5b58\u5728\u548c\u4f4d\u7f6e\u6bd4\u627e\u5230\u7279\u5b9a\u7c7b\u578b\u66f4\u91cd\u8981\uff09\u7b49\u3002<\/div>\n
\n
\n

\u672c\u6587\u7ecf\u8ba1\u7b97\u673a\u89c6\u89c9\u7814\u7a76\u9662\u516c\u4f17\u53f7\u6388\u6743\u8f6c\u8f7d\uff0c\u8f6c\u8f7d\u8bf7\u8054\u7cfb\u51fa\u5904\u3002<\/p>\n

\"\u56fe\u7247\"<\/p>\n

\u4e00\u3001\u7b80\u8981<\/h4>\n

\u76ee\u6807\u68c0\u6d4b\u6a21\u578b\u5728\u5b9a\u4f4d\u548c\u5206\u7c7b\u8bad\u7ec3\u671f\u95f4\u663e\u793a\u76ee\u6807\u65f6\u8868\u73b0\u826f\u597d\uff0c\u7136\u800c\uff0c\u7531\u4e8e\u521b\u5efa\u548c\u6ce8\u91ca\u68c0\u6d4b\u6570\u636e\u96c6\u7684\u96be\u5ea6\u548c\u6210\u672c\uff0c\u8bad\u7ec3\u8fc7\u7684\u6a21\u578b\u68c0\u6d4b\u5230\u6570\u91cf\u6709\u9650\u7684\u76ee\u6807\u7c7b\u578b\uff0c\u672a\u77e5\u76ee\u6807\u88ab\u89c6\u4e3a\u80cc\u666f\u5185\u5bb9\u3002\u8fd9\u963b\u788d\u4e86\u4f20\u7edf\u68c0\u6d4b\u5668\u5728\u73b0\u5b9e\u5e94\u7528\u4e2d\u7684\u91c7\u7528\uff0c\u5982\u5927\u89c4\u6a21\u7269\u4f53\u5339\u914d\u3001visual grounding\u3001\u89c6\u89c9\u5173\u7cfb\u9884\u6d4b\u3001\u969c\u788d\u68c0\u6d4b\uff08\u786e\u5b9a\u7269\u4f53\u7684\u5b58\u5728\u548c\u4f4d\u7f6e\u6bd4\u627e\u5230\u7279\u5b9a\u7c7b\u578b\u66f4\u91cd\u8981\uff09\u7b49\u3002<\/p>\n

\"\u56fe\u7247\"<\/p>\n

\u6709\u7814\u7a76\u8005\u63d0\u51fa\u7c7b\u4e0d\u53ef\u77e5\u76ee\u6807\u68c0\u6d4b\u4f5c\u4e3a\u4e00\u4e2a\u65b0\u95ee\u9898\uff0c\u4e13\u6ce8\u4e8e\u68c0\u6d4b\u5bf9\u8c61\u7684\u5bf9\u8c61\u7c7b\u3002\u5177\u4f53\u5730\u8bf4\uff0c\u5176\u76ee\u6807\u662f\u9884\u6d4b\u56fe\u50cf\u4e2d\u6240\u6709\u5bf9\u8c61\u7684\u8fb9\u754c\u6846\uff0c\u800c\u4e0d\u662f\u9884\u6d4b\u5b83\u4eec\u7684\u5bf9\u8c61\u7c7b\u3002\u9884\u6d4b\u7684\u6846\u53ef\u4ee5\u88ab\u53e6\u4e00\u4e2a\u7cfb\u7edf\u4f7f\u7528\uff0c\u4ee5\u6267\u884c\u7279\u5b9a\u4e8e\u5e94\u7528\u7a0b\u5e8f\u7684\u5206\u7c7b\u3001\u68c0\u7d22\u7b49\u3002<\/p>\n

\u63d0\u51fa\u4e86\u9488\u5bf9\u7c7b\u4e0d\u53ef\u77e5\u68c0\u6d4b\u5668\u7684\u57fa\u51c6\u6d4b\u8bd5\u7684\u8bad\u7ec3\u548c\u8bc4\u4f30\u534f\u8bae\uff0c\u4ee5\u63a8\u8fdb\u8be5\u9886\u57df\u7684\u672a\u6765\u7814\u7a76\u3002\u6700\u540e\uff0c\u7814\u7a76\u8005\u63d0\u51fa\u4e86\uff1a\uff081)\u57fa\u7ebf\u65b9\u6cd5\u548c(2\uff09\u4e00\u4e2a\u65b0\u7684\u7528\u4e8e\u7c7b\u65e0\u5173\u68c0\u6d4b\u7684\u5bf9\u6297\u6027\u5b66\u4e60\u6846\u67b6\uff0c\u5b83\u8feb\u4f7f\u6a21\u578b\u4ece\u7528\u4e8e\u9884\u6d4b\u7684\u7279\u5f81\u4e2d\u6392\u9664\u7279\u5b9a\u4e8e\u7c7b\u7684\u4fe1\u606f\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u5bf9\u6297\u6027\u5b66\u4e60\u65b9\u6cd5\u63d0\u9ad8\u4e86\u7c7b\u4e0d\u53ef\u77e5\u6027\u7684\u68c0\u6d4b\u6548\u7387\u3002<\/p>\n

\u4e8c\u3001\u80cc\u666f\u4ecb\u7ecd<\/h4>\n

\u4f1a\u6709\u540c\u5b66\u4f1a\u95ee\u201c\u76ee\u6807\u68c0\u6d4b\u4e2dbbox\u56de\u5f52\u4e2dclass-agnostic\u548cclass-specific\u7684\u533a\u522b\u5728\u54ea\uff1f\u201d\u3002\u4eca\u5929\u6211\u4eec\u7b80\u5355\u8bf4\u4e00\u4e0b\uff0c\u7136\u540e\u5f00\u59cb\u4eca\u5929\u4e3b\u6846\u67b6\u7684\u5206\u6790\u3002<\/p>\n

    \n
  • class-specific \u65b9\u5f0f\uff1a<\/strong>\u5f88\u591a\u5730\u65b9\u4e5f\u79f0\u4f5cclass-aware\u7684\u68c0\u6d4b\uff0c\u662f\u65e9\u671fFaster RCNN\u7b49\u4f17\u591a\u7b97\u6cd5\u91c7\u7528\u7684\u65b9\u5f0f\u3002\u5b83\u5229\u7528\u6bcf\u4e00\u4e2aRoI\u7279\u5f81\u56de\u5f52\u51fa\u6240\u6709\u7c7b\u522b\u7684bbox\u5750\u6807\uff0c\u6700\u540e\u6839\u636eclassification \u7ed3\u679c\u7d22\u5f15\u5230\u5bf9\u5e94\u7c7b\u522b\u7684box\u8f93\u51fa\u3002\u8fd9\u79cd\u65b9\u5f0f\u5bf9\u4e8ems coco\u670980\u7c7b\u524d\u666f\u7684\u6570\u636e\u96c6\u6765\u8bf4\uff0c\u5e76\u4e0d\u7b97\u6548\u7387\u9ad8\u7684\u505a\u6cd5\u3002<\/li>\n
  • class-agnostic \u65b9\u5f0f\uff1a<\/strong>\u53ea\u56de\u5f522\u7c7bbounding box\uff0c\u5373\u524d\u666f\u548c\u80cc\u666f\uff0c\u7ed3\u5408\u6bcf\u4e2abox\u5728classification \u7f51\u7edc\u4e2d\u5bf9\u5e94\u7740\u6240\u6709\u7c7b\u522b\u7684\u5f97\u5206\uff0c\u4ee5\u53ca\u68c0\u6d4b\u9608\u503c\u6761\u4ef6\uff0c\u5c31\u53ef\u4ee5\u5f97\u5230\u56fe\u7247\u4e2d\u6240\u6709\u7c7b\u522b\u7684\u68c0\u6d4b\u7ed3\u679c\u3002\u5f53\u7136\uff0c\u8fd9\u79cd\u65b9\u5f0f\u6700\u7ec8\u4e0d\u540c\u7c7b\u522b\u7684\u68c0\u6d4b\u7ed3\u679c\uff0c\u53ef\u80fd\u5305\u542b\u540c\u4e00\u4e2a\u524d\u666f\u6846\uff0c\u4f46\u5b9e\u9645\u5bf9\u7cbe\u5ea6\u7684\u5f71\u54cd\u4e0d\u7b97\u5f88\u5927\uff0c\u6700\u91cd\u8981\u7684\u662f\u5927\u5e45\u51cf\u5c11\u4e86bbox\u56de\u5f52\u53c2\u6570\u91cf\u3002\u5177\u4f53\u7ec6\u8282\uff0c\u81ea\u5df1\u53c2\u8003\u76ee\u524d\u4e00\u4e9b\u5f00\u6e90\u7b97\u6cd5\u6e90\u7801\u4f1a\u7406\u89e3\u7684\u66f4\u597d\u3002\uff08\u6458\u81ea\u4e8e\u77e5\u4e4e\u5305\u6587\u97ec\uff09<\/li>\n<\/ul>\n

    Class-agnostic\u76ee\u6807\u68c0\u6d4b\u5668\u4f7f\u7528object proposal methods (OPMs), conventional class-aware detectors\u548c\u63d0\u51fa\u7684adversarially trained class-agnostic detectors\u3002\u5982\u4e0b\u56fe\uff1a<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u7eff\u8272\u548c\u7d2b\u7ea2\u8272\u5206\u522b\u662f\u771f\u503c\u548c\u68c0\u6d4b\u7ed3\u679c\u3002\u4e09\u3001\u65b0\u6846\u67b6<\/strong><\/p>\n

      \n
    • General Framework<\/li>\n<\/ul>\n

      \u4f20\u7edf\u7684\u7c7b\u611f\u77e5\u68c0\u6d4b\u4fa7\u91cd\u4e8e\u68c0\u6d4b\u201c\u611f\u5174\u8da3\u7684\u5bf9\u8c61\u201d\uff0c\u8fd9\u672c\u8d28\u4e0a\u8981\u6c42\u6a21\u578b\u80fd\u591f\u533a\u5206\u5c01\u95ed\u5df2\u77e5\u96c6\u5408\u4e2d\u7684\u5bf9\u8c61\u7c7b\u578b\u3002\u76f4\u89c2\u5730\u8bf4\uff0c\u6a21\u578b\u901a\u8fc7\u7f16\u7801\u533a\u5206\u5bf9\u8c61\u7c7b\u578b\u7684\u7279\u5f81\u6765\u5b9e\u73b0\u8fd9\u4e00\u70b9\u3002\u7136\u800c\uff0c\u4e3a\u4e86\u4f7f\u7c7b\u4e0d\u53ef\u77e5\u7684\u68c0\u6d4b\u548c\u6a21\u578b\u80fd\u591f\u68c0\u6d4b\u5230\u4ee5\u524d\u770b\u4e0d\u89c1\u7684\u5bf9\u8c61\u7c7b\u578b\uff0c\u68c0\u6d4b\u5668\u5e94\u8be5\u7f16\u7801\u80fd\u591f\u66f4\u6709\u6548\u5730\u533a\u5206\u5bf9\u8c61\u4e0e\u80cc\u666f\u5185\u5bb9\u3001\u5355\u4e2a\u5bf9\u8c61\u4e0e\u56fe\u50cf\u4e2d\u7684\u5176\u4ed6\u5bf9\u8c61\u7684\u7279\u5f81\uff0c\u800c\u4e0d\u533a\u5206\u5bf9\u8c61\u7c7b\u578b\u3002<\/p>\n

      \"\u56fe\u7247\"<\/p>\n

      \u8bad\u7ec3\u4f20\u7edf\u7684\u76ee\u6807\u68c0\u6d4b\u5668\u7684\u4e8c\u5143\u5206\u7c7b\u4efb\u52a1\u4ee5\u53ca\u8fb9\u754c\u6846\u56de\u5f52\u4e0d\u8db3\u4ee5\u786e\u4fdd\u6a21\u578b\u5173\u6ce8\u7c7b\u65e0\u5173\u7279\u5f81\uff0c\u66f4\u91cd\u8981\u7684\u662f\uff0c\u5ffd\u7565\u7c7b\u578b\u533a\u5206\u7279\u5f81\uff0c\u4ee5\u4fbf\u66f4\u597d\u5730\u63a8\u5e7f\u5230\u770b\u4e0d\u89c1\u7684\u76ee\u6807\u7c7b\u578b\u3002\u4e3a\u4e86\u514b\u670d\u8fd9\u4e2a\u95ee\u9898\uff0c\u7814\u7a76\u8005\u5efa\u8bae\u4ee5\u4e00\u79cd\u5bf9\u6297\u6027\u7684\u65b9\u5f0f\u8bad\u7ec3\u7c7b\u4e0d\u53ef\u77e5\u7684\u76ee\u6807\u68c0\u6d4b\u5668\uff0c\u4ee5\u4fbf\u6a21\u578b\u56e0\u7f16\u7801\u5305\u542b\u76ee\u6807\u7c7b\u578b\u4fe1\u606f\u7684\u7f16\u7801\u7279\u5f81\u800c\u53d7\u5230\u60e9\u7f5a\u3002<\/p>\n

       <\/p>\n

      \u7814\u7a76\u8005\u63d0\u8bae\u7528\u5bf9\u6297\u6027\u9274\u522b\u5668\u5206\u652f\u6765\u589e\u5f3a\u7c7b\u4e0d\u53ef\u77e5\u7684\u68c0\u6d4b\u5668\uff0c\u8fd9\u4e9b\u5206\u652f\u8bd5\u56fe\u4ece\u68c0\u6d4b\u7f51\u7edc\u4e0a\u6e38\u8f93\u51fa\u7684\u7279\u5f81\u4e2d\u5206\u7c7b\u5bf9\u8c61\u7c7b\u578b\uff08\u5728\u8bad\u7ec3\u6570\u636e\u4e2d\u6ce8\u91ca\uff09\uff0c\u5982\u679c\u6a21\u578b\u8bad\u7ec3\u6210\u529f\uff0c\u5219\u5bf9\u5176\u8fdb\u884c\u60e9\u7f5a\u3002\u6a21\u578b\u4ee5\u4ea4\u66ff\u7684\u65b9\u5f0f\u8bad\u7ec3\uff0c\u8fd9\u6837\u5f53\u6a21\u578b\u7684\u5176\u4f59\u90e8\u5206\u66f4\u65b0\u65f6\uff0c\u9274\u522b\u5668\u88ab\u51bb\u7ed3\uff0c\u53cd\u4e4b\u4ea6\u7136\u3002\u5728\u66f4\u65b0\u9274\u522b\u5668\u65f6\uff0c\u7814\u7a76\u8005\u4f7f\u7528\u6807\u51c6\u7684\u5206\u7c7b\u4ea4\u53c9\u71b5\u635f\u5931\u7684\u76ee\u6807\u7c7b\u578b\u4f5c\u4e3a\u9884\u6d4b\u76ee\u6807\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u5728\u8bad\u7ec3\u6a21\u578b\u7684\u5176\u4f59\u90e8\u5206\u65f6\uff0c\u6700\u5c0f\u5316(a)\u76ee\u6807\u4e0e\u5426\u5206\u7c7b\u7684\u4ea4\u53c9\u71b5\u635f\u5931\uff0c(b)\u8fb9\u754c\u6846\u56de\u5f52\u7684\u5e73\u6ed1L1\u635f\u5931\uff0c\u4ee5\u53ca(c)\u9274\u522b\u5668\u9884\u6d4b\u7684\u8d1f\u71b5\u3002\u8fd9\u79cd\u71b5\u6700\u5927\u5316\u8feb\u4f7f\u68c0\u6d4b\u6a21\u578b\u7684\u4e0a\u6e38\u90e8\u5206\u4ece\u5176\u8f93\u51fa\u7684\u7279\u5f81\u4e2d\u6392\u9664\u76ee\u6807\u7c7b\u578b\u4fe1\u606f\u3002\u5bf9\u4e8e\u6a21\u578b\u7684\u6bcf\u6b21\u66f4\u65b0\uff0c\u9274\u522b\u5668\u88ab\u66f4\u65b0\u4e94\u6b21\uff0c\u5728\u6574\u4e2a\u76ee\u6807\u4e2d\u4f7f\u7528\u4e58\u5b50\u03b1\uff08\u8c03\u6574{0.1,1}\uff09\u5bf9\u8d1f\u71b5\u8fdb\u884c\u52a0\u6743\u3002\u4e0a\u56fe\u603b\u7ed3\u4e86\u5b8c\u6574\u7684\u6846\u67b6\u3002<\/p>\n

      \u56db\u3001\u5b9e\u9a8c<\/h4>\n

      \"\u56fe\u7247\"<\/p>\n

      Generalization results for FRCNN models trained on the seen VOC dataset. The top row shows macro-level AR@kfor seen and unseen classes in VOC and their harmonic mean (AR-HM). FRCNN-agnostic-adv performs the best overall. The second row shows micro-level results for the easy, medium, and hard unseen classes. FRCNN-agnostic-adv performs the best on the hard and easy classes with recall drop for the medium class. The last row provides results of evaluation on the COCO data of 60 unseen classes. FRCNN-agnostic-adv achieves the best AR@k\u00a0for objects of all sizes.<\/p>\n

      \"\u56fe\u7247\"<\/p>\n

      \"\u56fe\u7247\"<\/p>\n

      Generalization results for SSD models trained on the seen VOC dataset. The top row shows macro-level AR@kfor seen and unseen classes in VOC as well as their harmonic mean (AR-HM). SSD-agnostic-adv performs the best on AR- Unseen and AR-HM, with a drop in AR-Seen, but the models that outperform SSD-agnostic-adv on AR-Seen do significantly worse on AR-Unseen and AR-HM. The second row shows micro-level results for the easy, medium, and hard unseen classes. SSD-agnostic-adv performs the best in all categories. The last row provides results of evaluation on the COCO data of 60 unseen classes. SSD-agnostic-adv achieves the best AR@k with a slight reduction for small-sized objects.<\/p>\n

      \"\u56fe\u7247\"<\/p>\n<\/div>\n<\/div>\n

      <\/div>\n

      <\/g><\/svg><\/i> \"Loading\"<\/p>\n

      <\/div>\n","protected":false},"excerpt":{"rendered":"

      \u76ee\u6807\u68c0\u6d4b\u6a21\u578b\u5728\u5b9a\u4f4d\u548c\u5206\u7c7b\u8bad\u7ec3\u671f\u95f4\u663e\u793a\u76ee\u6807\u65f6\u8868\u73b0\u826f\u597d\uff0c\u7136\u800c\uff0c\u7531\u4e8e\u521b\u5efa\u548c\u6ce8\u91ca\u68c0\u6d4b\u6570\u636e\u96c6\u7684\u96be\u5ea6\u548c\u6210\u672c\uff0c\u8bad\u7ec3\u8fc7\u7684\u6a21\u578b\u68c0 […]<\/p>\n

      <\/div>\n

      <\/g><\/svg><\/i> \"Loading\"<\/p>\n

      <\/div>\n","protected":false},"author":56,"featured_media":21597,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[23,20,80],"tags":[1244],"class_list":["post-21595","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-23","category-20","category-80","tag-1244"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/21595","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/56"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=21595"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/21595\/revisions"}],"predecessor-version":[{"id":21598,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/21595\/revisions\/21598"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/21597"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=21595"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=21595"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=21595"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}