{"id":7128,"date":"2023-12-26T15:47:39","date_gmt":"2023-12-26T07:47:39","guid":{"rendered":"http:\/\/192.168.10.115\/?p=7128"},"modified":"2023-12-26T15:47:39","modified_gmt":"2023-12-26T07:47:39","slug":"2023-12-26-%e9%ab%98%e7%b2%be%e5%9c%b0%e5%9b%be%e6%96%b0sota%ef%bc%81pivotnet%ef%bc%9a%e7%ab%af%e5%88%b0%e7%ab%af%e5%90%91%e9%87%8f%e5%8c%96%e9%ab%98%e7%b2%be%e5%9c%b0%e5%9b%be%e6%9e%84%e5%bb%ba","status":"publish","type":"post","link":"http:\/\/222.128.65.89:18086\/index.php\/2023\/12\/26\/7128\/","title":{"rendered":"2023-12-26 \u9ad8\u7cbe\u5730\u56fe\u65b0SOTA\uff01PivotNet\uff1a\u7aef\u5230\u7aef\u5411\u91cf\u5316\u9ad8\u7cbe\u5730\u56fe\u6784\u5efa\u65b9\u6848\uff08ICCV2023\uff09"},"content":{"rendered":"\n
\"\"<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2308.16477<\/a><\/p>\n\n\n\n

\u9ad8\u6e05\u5730\u56fe\uff08HD map\uff09\u662f\u81ea\u52a8\u9a7e\u9a76\u7814\u7a76\u4e2d\u6700\u5173\u952e\u7684\u7ec4\u6210\u90e8\u5206\u4e4b\u4e00\u3002\u4f20\u7edf\u7684HD\u5730\u56fe\u6784\u5efa\u4f9d\u8d56\u4e8e\u5bf9\u6fc0\u5149\u96f7\u8fbe\u70b9\u4e91\u8fdb\u884c\u624b\u52a8\u6807\u6ce8\uff0c\u8017\u65f6\u4e14\u8d39\u529b\u3002\u8fd1\u671f\u7684\u7814\u7a76\u901a\u8fc7\u5730\u56fe\u5b66\u4e60 (map learning) \u95ee\u9898\u6765\u51cf\u5c11\u6807\u6ce8\u6210\u672c\u3002\u7ed9\u5b9a\u8f66\u8f7d\u4f20\u611f\u5668\u6570\u636e\uff0c\u5730\u56fe\u5b66\u4e60\u7684\u76ee\u6807\u662f\u5728\u9884\u5b9a\u4e49\u7684\u9e1f\u77b0\u89c6\u56fe\uff08BEV\uff09\u8303\u56f4\u5185\u6784\u5efa\u5c40\u90e8\u5730\u56fe\u3002\u5927\u591a\u6570\u73b0\u6709\u65b9\u6cd5\u5c06\u5730\u56fe\u6784\u5efa\u89c6\u4e3a\u4e00\u79cd\u8bed\u4e49\u5b66\u4e60 (semantic learning) \u95ee\u9898\u3002\u5b83\u4eec\u5c06\u5730\u56fe\u8868\u793a\u4e3a\u4e00\u7ec4\u56fa\u5b9a\u6570\u91cf\u7684\u7f51\u683c\uff0c\u5e76\u4e3a\u6bcf\u4e2a\u7f51\u683c\u9884\u6d4b\u7c7b\u522b\u6807\u7b7e, \u6700\u7ec8\u901a\u8fc7\u590d\u6742\u7684\u540e\u5904\u7406\u751f\u6210\u4e0b\u6e38\u6240\u9700\u7684\u5411\u91cf\u5316\u5730\u56fe\u3002\u8fd9\u79cd\u8bed\u4e49\u8868\u793a\u5b58\u5728\u660e\u663e\u7684\u5c40\u9650\u6027\uff0c\u5305\u62ec\u5197\u4f59\u4fe1\u606f\u3001\u5ffd\u7565\u7f51\u683c\u4e4b\u95f4\u548c\u5185\u90e8\u7684\u51e0\u4f55\u5173\u7cfb\u4ee5\u53ca\u590d\u6742\u540e\u5904\u7406\u4e2d\u7684\u65f6\u95f4\u6d88\u8017\u548c\u7d2f\u8ba1\u8bef\u5dee\u3002<\/p>\n\n\n\n

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

\u4e3a\u4e86\u89e3\u51b3\u5f53\u524d\u8bed\u4e49\u5b66\u4e60\u65b9\u6cd5\u7684\u95ee\u9898\uff0c\u4e00\u4e9b\u65b9\u6cd5\u63d0\u51fa\u4ee5\u7aef\u5230\u7aef\u7684\u65b9\u5f0f\u751f\u6210\u77e2\u91cf\u5316\u8868\u793a\u3002MapTR \u4f7f\u7528\u56fa\u5b9a\u6570\u91cf\u7684\u70b9\u6765\u8868\u793a\u4e0d\u540c\u590d\u6742\u5ea6\u7684\u5730\u56fe\u5143\u7d20\u3002\u8fd9\u79cd\u65b9\u6cd5\u5b58\u5728\u7684\u95ee\u9898\u662f\uff1a\u9996\u5148\uff0c\u5747\u5300\u5316\u7684\u8868\u793a\u5305\u542b\u4e86\u5bf9\u51e0\u4f55\u6027\u5f71\u54cd\u8f83\u5c0f\u7684\u5197\u4f59\u70b9\u3002\u5176\u6b21\uff0c\u4f7f\u7528\u56fa\u5b9a\u6570\u91cf\u7684\u70b9\u8868\u793a\u52a8\u6001\u5f62\u72b6\u7684\u7ebf\u53ef\u80fd\u4f1a\u4e22\u5931\u5730\u56fe\u5143\u7d20\u7684\u5173\u952e\u7ec6\u8282\uff0c\u7279\u522b\u662f\u5bf9\u4e8e\u5706\u89d2\u548c\u76f4\u89d2\uff08\u56fe1\uff09\u3002\u56e0\u6b64\uff0c\u4e3a\u4e86\u5b66\u4e60\u51c6\u786e\u4e14\u7d27\u51d1\u7684\u8868\u793a\uff0c\u6211\u4eec\u5c06\u5730\u56fe\u5143\u7d20\u5efa\u6a21\u4e3a\u6709\u5e8f\u7684\u67a2\u7ebd\u70b9\u5217\u8868\uff0c\u8fd9\u79cd\u8868\u793a\u65b9\u6cd5\u7d27\u51d1\u3001\u9002\u5408\u5904\u7406\u89d2\u70b9\u4e14\u5177\u6709\u51e0\u4f55\u9c81\u68d2\u6027\u3002\u7136\u800c\uff0c\u7531\u4e8e\u4e0d\u540c\u5730\u56fe\u5143\u7d20\u5185\u67a2\u7ebd\u70b9\u7684\u6570\u91cf\u52a8\u6001\u53d8\u5316\uff0c\u8fd9\u7ed9\u57fa\u4e8e\u67a2\u7ebd\u70b9\u7684\u8868\u793a\u65b9\u5f0f\u5e26\u6765\u4e86\u65b0\u7684\u6311\u6218\u3002\u524d\u4eba\u7684\u5de5\u4f5c\u5229\u7528\u7c97\u5230\u7ec6\u7684\u6846\u67b6\u548c\u81ea\u56de\u5f52\u89e3\u7801\u6765\u89e3\u51b3\u8fd9\u4e9b\u6311\u6218\uff0c\u4f46\u81ea\u56de\u5f52\u7684\u7279\u6027\u4f1a\u5bfc\u81f4\u63a8\u7406\u65f6\u95f4\u8f83\u957f\u4e14\u4f1a\u7d2f\u79ef\u8bef\u5dee\u3002<\/p>\n\n\n\n

\u9488\u5bf9\u8fd9\u4e9b\u95ee\u9898\uff0c\u6211\u4eec\u63d0\u51fa\u4e86 PivotNet\uff0c\u5b83\u91c7\u7528\u4e86 set prediction \u7684\u8303\u5f0f\uff0c\u4f7f\u7528\u7edf\u4e00\u7684\u57fa\u4e8e\u67a2\u7ebd\u70b9\u7684\u5730\u56fe\u8868\u793a\u5bf9\u5730\u56fe\u5143\u7d20\u8fdb\u884c\u51c6\u786e\u5efa\u6a21\u3002<\/p>\n\n\n\n

\u672c\u6587\u7684\u8d21\u732e\u6709\u4e09\u4e2a\u65b9\u9762\uff1a<\/p>\n\n\n\n

    \n
  • \u6211\u4eec\u63d0\u51fa\u4e86PivotNet\uff0c\u57fa\u4e8e\u67a2\u7ebd\u70b9\u8868\u793a\uff0c\u5b9e\u73b0\u4e86\u7aef\u5230\u7aef\u7684\u7cbe\u786e\u800c\u7d27\u51d1\u7684\u77e2\u91cf\u5316\u5efa\u56fe\u3002
    <\/li>\n\n\n\n
  • \u6211\u4eec\u521b\u65b0\u6027\u5730\u5f15\u5165\u4e86\u70b9\u7ebf\u63a9\u7801\u6a21\u5757 (point-to-line mask module)\u3001\u67a2\u7ebd\u70b9\u52a8\u6001\u5339\u914d\u6a21\u5757 (pivot dynamic matching module) \u548c\u52a8\u6001\u77e2\u91cf\u5316\u5e8f\u5217\u635f\u5931 (dynamic vectorized sequence loss)\uff0c\u4ee5\u5b9e\u73b0\u51c6\u786e\u7684\u5730\u56fe\u5143\u7d20\u5efa\u6a21\u3002
    <\/li>\n\n\n\n
  • PivotNet \u5728\u5e7f\u6cdb\u4f7f\u7528\u7684\u516c\u5f00\u6570\u636e\u96c6\u4e2d\u8fdb\u884c\u4e86\u9a8c\u8bc1, \u5206\u522b\u5728 nuScenes \u53ca Argoverse 2 \u6570\u636e\u96c6\u4e0a\u9886\u5148\u5f53\u524d SOTA 11.2 \u548c 5.9 mAP\u3002
    <\/li>\n<\/ul>\n\n\n\n

    Pipeline<\/strong><\/h2>\n\n\n\n
    \"\"<\/figure>\n\n\n\n

    PivotNet\u7684\u6846\u67b6\u5982\u56fe2\u6240\u793a\uff0c\u5305\u62ec\u56db\u4e2a\u4e3b\u8981\u6a21\u5757\uff1a\u76f8\u673a\u7279\u5f81\u63d0\u53d6\u5668\uff08camera feature extractor\uff09\u3001BEV \u7279\u5f81\u89e3\u7801\u5668 (bev feature decoder)\u3001\u7ebf\u611f\u77e5\u7684\u70b9\u89e3\u7801\u5668 (line-aware point decoder) \u548c\u67a2\u7ebd\u70b9\u9884\u6d4b\u5668 (pivot point head)\u3002\u591a\u89c6\u56fe\u56fe\u50cf\u9996\u5148\u8f93\u5165\u76f8\u673a\u7279\u5f81\u63d0\u53d6\u5668\uff0c\u751f\u6210\u56fe\u50cf\u7279\u5f81\u3002\u7136\u540e\uff0c\u901a\u8fc7BEV \u7279\u5f81\u89e3\u7801\u5668\uff0c\u5c06\u56fe\u50cf\u7279\u5f81\u805a\u5408\u5e76\u8f6c\u6362\u4e3a\u7edf\u4e00\u7684 BEV \u7279\u5f81\u3002\u63a5\u7740\u7ebf\u611f\u77e5\u7684\u70b9\u89e3\u7801\u5668\u63d0\u53d6\u7ebf\u611f\u77e5\u70b9\u7279\u5f81\u3002\u6700\u540e\uff0c\u67a2\u7ebd\u70b9\u9884\u6d4b\u5668\u5220\u9664\u5171\u7ebf\u70b9\u5e76\u8f93\u51fa\u7075\u6d3b\u800c\u7d27\u51d1\u7684\u57fa\u4e8e\u67a2\u7ebd\u70b9\u7684\u5730\u56fe\u8868\u793a\u3002<\/p>\n\n\n\n

      \n
    1. \u76f8\u673a\u7279\u5f81\u63d0\u53d6\u5668\uff1a\u5c06\u8f66\u8f7d\u76f8\u673a\u6355\u83b7\u7684\u591a\u89c6\u89d2\u56fe\u50cf\u8f93\u5165\u5230\u76f8\u673a\u7279\u5f81\u63d0\u53d6\u5668\uff0c\u751f\u6210\u76f8\u673a\u89c6\u89d2\u7279\u5f81\u3002
      <\/li>\n\n\n\n
    2. BEV\u7279\u5f81\u89e3\u7801\u5668\uff1a\u91c7\u7528 Transformer-based \u6a21\u5757\u5c06\u56fe\u50cf\u7279\u5f81\u805a\u5408\u548c\u8f6c\u6362\u6210\u7edf\u4e00\u7684 BEV \u8868\u793a\uff0c\u901a\u8fc7\u91c7\u7528\u53ef\u53d8\u5f62\u6ce8\u610f\u529b\u673a\u5236\u6765\u7f16\u7801\u51e0\u4f55\u5bf9\u5e94\u5173\u7cfb\uff0c\u540c\u65f6\u51cf\u5c11\u8ba1\u7b97\u5185\u5b58\u3002\u8fd9\u4e2a\u6a21\u5757\u5c06\u56fe\u50cf\u7279\u5f81\u8f6c\u5316\u4e3aBEV\u7279\u5f81\u3002
      <\/li>\n\n\n\n
    3. \u7ebf\u611f\u77e5\u70b9\u89e3\u7801\u5668\uff1a\u901a\u8fc7\u7ebf\u611f\u77e5\u70b9\u89e3\u7801\u5668\uff0c\u5c06BEV\u7279\u5f81\u89e3\u7801\u6210\u7ebf\u611f\u77e5\u7684\u70b9\u7279\u5f81\u3002\u8be5\u6a21\u5757\u91c7\u7528Transformer-based\u67b6\u6784\uff0c\u5e76\u4e14\u901a\u8fc7\u5f15\u5165\u70b9\u5230\u7ebf\u63a9\u7801\u4ee5\u53ca\u67a2\u7ebd\u70b9\u52a8\u6001\u5339\u914d\u7b49\u6a21\u5757\uff0c\u80fd\u591f\u66f4\u51c6\u786e\u5730\u5bf9\u5730\u56fe\u5143\u7d20\u8fdb\u884c\u5efa\u6a21\u3002
      <\/li>\n\n\n\n
    4. \u67a2\u7ebd\u70b9\u9884\u6d4b\u5668\uff1a\u6700\u540e\uff0c\u67a2\u7ebd\u70b9\u9884\u6d4b\u5668\u6839\u636e\u7ebf\u611f\u77e5\u7684\u70b9\u7279\u5f81\uff0c\u9884\u6d4b\u51fa\u5730\u56fe\u4e2d\u7684\u67a2\u7ebd\u70b9\u3002\u8fd9\u4e9b\u67a2\u7ebd\u70b9\u7528\u4e8e\u5bf9\u5730\u56fe\u5143\u7d20\u8fdb\u884c\u5efa\u6a21\u3002
      <\/li>\n<\/ol>\n\n\n\n

      Point-Line Relation Modeling<\/strong><\/h2>\n\n\n\n
      \"\"<\/figure>\n\n\n\n

      \u5728 set prediction \u7684\u6846\u67b6\u4e0b\uff0c\u6211\u4eec\u65e0\u6cd5\u660e\u786e\u533a\u5206\u5b9e\u4f8b\u4e4b\u95f4\u548c\u5b9e\u4f8b\u5185\u90e8\u7684\u70b9\uff0c\u8fd9\u53ef\u80fd\u5bfc\u81f4\u5b9e\u4f8b\u6df7\u6dc6\u3002\u56e0\u6b64\uff0c\u4e3a\u4e86\u786e\u4fdd\u5730\u56fe\u5143\u7d20\u7684\u51c6\u786e\u5efa\u6a21\uff0c\u6211\u4eec\u9700\u8981\u5bf9\u7f51\u7edc\u4e2d\u7684\u70b9\u7ebf\u5173\u7cfb\u8fdb\u884c\u7f16\u7801\u3002\u70b9\u7ebf\u5173\u7cfb\u4e3b\u8981\u5305\u542b\u4ece\u5c5e (subordinate) \u5173\u7cfb\u548c\u51e0\u4f55 (geometrical) \u5173\u7cfb\u3002\u5176\u4e2d\uff0c\u4ece\u5c5e\u5173\u7cfb\u6307\u7684\u662f\u7f51\u7edc\u5b9a\u4e49\u7684\u70b9\u7ebf\u5bf9\u5e94\u5173\u7cfb\uff0c\u6bd4\u5982\u54ea\u4e9b point query \u5c5e\u4e8e\u540c\u4e00\u6761\u7ebf\uff0c\u54ea\u4e9b\u5c5e\u4e8e\u4e0d\u540c\u7684\u7ebf\u3002\u51e0\u4f55\u5173\u7cfb\u6307\u7684\u662f\u70b9\u7ebf\u7684\u51e0\u4f55\u5bf9\u5e94\uff0c\u6bd4\u5982\u70b9\u6784\u6210\u7ebf\uff0c\u4e00\u6761\u7ebf\u5bf9\u5e94\u7684\u70b9\u5e8f\u5217\u5305\u542b\u4e86\u6062\u590d\u8fd9\u6761\u7ebf\u7684\u4efb\u610f\u8868\u8fbe\u6240\u9700\u8981\u7684\u6240\u6709\u4fe1\u606f\u3002<\/p>\n\n\n\n

      \u4e3a\u4e86\u7f16\u7801\u70b9\u7ebf\u5173\u7cfb\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u70b9\u7ebf\u63a9\u7801\u6a21\u5757<\/strong> (point-to-line mask module)\uff0c\u5b83\u4ee5\u76f4\u89c2\u7684\u65b9\u5f0f\u7f16\u7801\u4e86\u4ece\u5c5e\u5173\u7cfb\u548c\u51e0\u4f55\u5173\u7cfb\u3002\u8be5\u6a21\u5757\u7684\u57fa\u672c\u6982\u5ff5\u662f\u4e00\u4e2a\u5411\u91cf\u5316\u7684\u6709\u5e8f\u70b9\u5e8f\u5217\u80fd\u591f\u6784\u5efa\u4e00\u4e2a\u5730\u56fe\u5143\u7d20\u3002\u57fa\u4e8e\u8fd9\u4e2a\u6982\u5ff5\uff0c\u6211\u4eec\u5f3a\u5236\u8981\u6c42\u76f8\u540c\u5b9e\u4f8b\u7684\u70b9\u67e5\u8be2\uff08point query\uff09\u5b66\u4e60\u5171\u4eab\u7684\u7ebf\u611f\u77e5\u6ce8\u610f\u529b\u63a9\u7801 (line-aware mask)\u3002\u8be5\u63a9\u7801\u4f1a\u53c2\u4e0e transformer \u7684\u4ea4\u53c9\u6ce8\u610f\u529b\u8ba1\u7b97\uff0c\u4f7f\u5f97\u70b9\u67e5\u8be2\u7684\u6ce8\u610f\u529b\u88ab\u9650\u5236\u5728\u5bf9\u5e94\u7684\u63a9\u7801\u533a\u57df\u5185\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u540c\u4e00\u5b9e\u4f8b\u7684\u70b9\u67e5\u8be2\uff08point query\uff09\u5171\u4eab\u540c\u6837\u7684\u63a9\u7801, \u4e0d\u540c\u5b9e\u4f8b\u91c7\u7528\u4e0d\u540c\u7684\u63a9\u7801\u3002\u6bcf\u4e2a\u5b9e\u4f8b\u90fd\u6709\u4e00\u4e2a\u552f\u4e00\u7684\u63a9\u7801\u8d1f\u8d23\u5904\u7406\u8be5\u5b9e\u4f8b\u7684\u6240\u6709\u70b9\u67e5\u8be2\u3002\u901a\u8fc7\u9650\u5236\u6ce8\u610f\u529b\u533a\u57df\uff0c\u70b9\u7ebf\u63a9\u7801\u6a21\u5757\u4ece\u7ed3\u6784\u4e0a\u5bf9\u4ece\u5c5e\u5173\u7cfb\u8fdb\u884c\u4e86\u7f16\u7801\uff0c\u4f7f\u5f97\u70b9\u7279\u5f81\u80fd\u66f4\u597d\u5730\u6355\u6349\u5230\u5730\u56fe\u5143\u7d20\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u6b64\u5916\uff0c\u901a\u8fc7\u76d1\u7763\u7ebf\u611f\u77e5\u63a9\u7801 \uff08line-aware mask\uff09\uff0c\u8ba9\u5b83\u53bb\u5b66\u4e60\u8fd9\u6761\u7ebf\u7684\u8bed\u4e49\u8868\u8fbe\uff0c\u70b9\u7ebf\u7684\u51e0\u4f55\u5173\u7cfb\u4e5f\u5f97\u5230\u4e86\u663e\u5f0f\u7ea6\u675f\u3002<\/p>\n\n\n\n

      Pivot Dynamic Matching<\/strong><\/h2>\n\n\n\n
      \"\"<\/figure>\n\n\n\n

      \u7528\u67a2\u7ebd\u70b9\u5e8f\u5217\u5bf9\u5730\u56fe\u5143\u7d20\u8fdb\u884c\u5efa\u6a21\u7684\u96be\u70b9\u5728\u4e8e\uff1a<\/p>\n\n\n\n

        \n
      1. \u67a2\u7ebd\u70b9\u4e4b\u95f4\u7684\u62d3\u6251\u5173\u7cfb\u5efa\u6a21\uff0c\u5373\u9700\u8981\u4fdd\u8bc1\u70b9\u5e8f\u5217\u7684\u6709\u5e8f\u6027<\/li>\n\n\n\n
      2. \u52a8\u6001\u67a2\u7ebd\u70b9\u6570\u91cf\u5efa\u6a21\uff0c\u5373\u9700\u8981\u6839\u636e\u5730\u56fe\u5143\u7d20\u7684\u590d\u6742\u5ea6\u52a8\u6001\u8c03\u6574\u67a2\u7ebd\u70b9\u6570\u91cf\u3002<\/li>\n<\/ol>\n\n\n\n

        Set prediction \u4efb\u52a1\u5e38\u7528\u7684\u4e8c\u5206\u5339\u914d\uff08bipartite matching\uff09\u65e0\u6cd5\u4fdd\u8bc1\u5339\u914d\u5143\u7d20\u95f4\u7684\u6709\u5e8f\u6027\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u5f15\u5165\u4e86 \u5e8f\u5217\u5339\u914d\uff08sequence matching\uff09\u7684\u6982\u5ff5\uff0c\u5e76\u8bbe\u8ba1\u4e86\u67a2\u7ebd\u70b9\u52a8\u6001\u5339\u914d\u6a21\u5757<\/strong> (pivot dynamic matching module)\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5728 PivotNet \u7684\u6846\u67b6\u4e0b\uff0c\u4e0d\u540c\u5730\u56fe\u5143\u7d20\u4e4b\u95f4\uff08line level\uff09\u662f\u65e0\u5e8f\u7684\uff0c\u4f46\u5730\u56fe\u5143\u7d20\u5185\u90e8\u7684\u70b9 (point level) \u662f\u6709\u5e8f\u7684\u3002<\/p>\n\n\n\n

        \u6211\u4eec\u628a\u5730\u56fe\u5143\u7d20\u5efa\u6a21\u6210\u4e00\u4e2a\u540c\u65f6\u5305\u542b\u67a2\u7ebd\u70b9\uff08pivot point\uff09\u548c\u5171\u7ebf\u70b9\uff08collinear point\uff09\u7684\u5e8f\u5217\u3002\u67a2\u7ebd\u70b9\u6307\u7684\u662f\u5bf9\u5730\u56fe\u5143\u7d20\u5f62\u72b6\u8d21\u732e\u5ea6\u5927\u7684\u70b9\uff0c\u800c\u5171\u7ebf\u70b9\u5219\u6307\u4ee3\u76f8\u90bb\u4e24\u4e2a\u67a2\u7ebd\u70b9\u4e4b\u95f4\u7684\u70b9\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u8bbe\u5b9a\u4e86\u4e00\u4e2a\u5730\u56fe\u5143\u7d20\u7684\u6700\u5927\u70b9\u6570\u91cf N, \u7528\u8fd9\u4e2a\u957f\u5ea6\u4e3a N \u7684\u70b9\u5e8f\u5217\u6765\u4ee3\u8868\u4e00\u4e2a\u5b9e\u4f8b\uff0c\u8fd9\u4e2a\u70b9\u5e8f\u5217\u540c\u65f6\u5305\u542b\u4e86\u67a2\u7ebd\u70b9\u548c\u5171\u7ebf\u70b9\u3002\u7136\u540e\u6211\u4eec\u7528\u8fd9\u4e2a\u56fa\u5b9a\u957f\u5ea6\u7684 DT \u70b9\u5e8f\u5217\u548c\u52a8\u6001\u957f\u5ea6\uff08T\uff09\u7684 GT \u67a2\u7ebd\u70b9\u5e8f\u5217\u53bb\u505a pivot dynamic matching\u3002\u5339\u914d\u4e0a\u7684\u70b9\u6211\u4eec\u8ba4\u4e3a\u662f\u67a2\u7ebd\u70b9\uff0c\u672a\u5339\u914d\u4e0a\u7684\u8ba4\u4e3a\u662f\u5171\u7ebf\u70b9\uff0c\u540c\u65f6\u6211\u4eec\u4f1a\u6839\u636e\u5339\u914d\u7ed3\u679c\u5bf9\u8fd9\u4e9b\u70b9\u8fdb\u884c\u5206\u7c7b\u3002\u5728\u6d4b\u8bd5\u9636\u6bb5\uff0c\u88ab\u9884\u6d4b\u4e3a\u5171\u7ebf\u70b9\u7684\u70b9\u4f1a\u88ab\u5220\u6389\uff0c\u53ea\u4fdd\u7559\u5e8f\u5217\u4e2d\u7684\u67a2\u7ebd\u70b9\u4f5c\u4e3a\u6700\u7ec8\u8f93\u51fa\u3002<\/p>\n\n\n\n

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

        Dynamic Vectorized Sequence Loss<\/strong><\/h2>\n\n\n\n

        \u52a8\u6001\u77e2\u91cf\u5316\u5e8f\u5217\u635f\u5931\uff08Dynamic Vectorized Sequence loss\uff0c\u7b80\u79f0DVS loss\uff09\u76d1\u7763\u4e86\u77e2\u91cf\u5316\u5730\u56fe\u5143\u7d20\u9884\u6d4b\u4e2d\u4f4d\u7f6e\u548c\u62d3\u6251\u5173\u7cfb\u3002\u5b83\u5305\u542b\u4e09\u4e2a\u4e3b\u8981\u90e8\u5206\uff1a\u67a2\u7ebd\u70b9\u76d1\u7763\u3001\u5171\u7ebf\u70b9\u76d1\u7763\u548c\u67a2\u7ebd\u70b9\u5206\u7c7b\u635f\u5931\u3002<\/p>\n\n\n\n

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

        \u603b\u7684\u6765\u8bf4\uff0c\u52a8\u6001\u77e2\u91cf\u5316\u5e8f\u5217\u635f\u5931\u7ed3\u5408\u4e86\u67a2\u7ebd\u70b9\u3001\u5171\u7ebf\u70b9\u548c\u62d3\u6251\u5173\u7cfb\u7684\u76d1\u7763\uff0c\u901a\u8fc7\u6743\u91cd\u56e0\u5b50\u8fdb\u884c\u7efc\u5408\u8ba1\u7b97\uff0c\u7528\u4e8e\u6307\u5bfc\u77e2\u91cf\u5316\u5730\u56fe\u5143\u7d20\u7684\u9884\u6d4b\u548c\u4f18\u5316\u3002<\/p>\n\n\n\n

        \u5b9e\u9a8c<\/strong><\/h2>\n\n\n\n

        Comparison with SOTA<\/strong><\/h3>\n\n\n\n

        \u5728nuScenes\u6570\u636e\u96c6\u4e0a\uff0c\u6211\u4eec\u5c06PivotNet\u4e0e\u5f53\u524d SOTA \u8fdb\u884c\u4e86\u6bd4\u8f83\u3002\u88681\u4e2d\u5c55\u793a\u4e86\u4e0d\u540c\u5b9e\u9a8c\u8bbe\u7f6e\u4e0b PivotNet \u4e0e\u73b0\u6709\u65b9\u6cd5\u7684\u6574\u4f53\u8bc4\u4f30\u6027\u80fd\u5bf9\u6bd4\u7ed3\u679c\u3002\u73b0\u6709\u65b9\u6cd5\u5728\u8bc4\u4f30\u4e2d\u4f7f\u7528\u4e86\u4e0d\u540c\u7684 Chamfer Distance \u9608\u503c\u6765\u8ba1\u7b97 AP\u3002\u56e0\u6b64\uff0c\u4e3a\u4e86\u8fdb\u884c\u516c\u5e73\u6bd4\u8f83\uff0c\u6211\u4eec\u5728\u88681\u4e2d\u5bf9 PivotNet \u5206\u522b\u5728\u4e24\u79cd\u9608\u503c\u8bbe\u7f6e\u4e0b\u8fdb\u884c\u4e86\u8bc4\u4f30\u3002\u4e0e MapTR \u76f8\u6bd4\uff0c\u5728 {0.2, 0.5, 1.0}m \u7684\u8bc4\u6d4b\u9608\u503c\u4e0b\uff0c\u6211\u4eec\u9886\u5148\u4e86 11.2% mAP, \u5728 {0.5,1.0,1.5}m \u7684\u8bc4\u6d4b\u4e0e\u4e4b\u4e0b\u9886\u5148 7.3% mAP\u3002<\/p>\n\n\n\n

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

        \u88682\u5bf9PivotNet\u5728\u65b0\u7684\u5927\u89c4\u6a21\u6570\u636e\u96c6Argoverse 2\u4e0a\u8fdb\u884c\u4e86\u8bc4\u4f30\uff0c\u5e76\u4e0eSOTA\u65b9\u6cd5\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002\u6211\u4eec\u4f7f\u7528\u516c\u5f00\u7684\u6e90\u4ee3\u7801\u590d\u73b0\u4e86\u8fd9\u4e9b\u65b9\u6cd5\uff0c\u5e76\u5c06\u5176\u8fc1\u79fb\u5230\u4e86 Argoverse 2 \u6570\u636e\u96c6\u4e0a\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u88682\u4e2d\u7684\u6240\u6709\u7ed3\u679c\u90fd\u662f\u4f7f\u7528{0.2\uff0c0.5\uff0c1.0}m\u9608\u503c\u8fdb\u884c\u8bc4\u4f30\u7684\u3002\u5982\u8868\u6240\u793a\uff0cPivotNet \u5728 Argoverse 2 \u9886\u5148\u73b0\u6709 SOTA 5.9% mAP\u3002<\/p>\n\n\n\n

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

        Ablation Study<\/strong><\/h3>\n\n\n\n
        \"\"<\/figure>\n\n\n\n

        \u6211\u4eec\u5728\u88683\u4e2d\u4ed4\u7ec6\u5206\u6790\u4e86\u6bcf\u4e2a\u6a21\u5757\u5bf9PivotNet\u6700\u7ec8\u6027\u80fd\u7684\u8d21\u732e\u3002\u7b2c\u4e00\u884c\u4ee3\u8868\u57fa\u51c6\u65b9\u6cd5\uff0c\u5b83\u91c7\u7528\u4e86\u57fa\u4e8e transformer \u7684\u666e\u901a\u70b9\u89e3\u7801\u5668\u548c\u57fa\u4e8e\u5206\u7c7b\u7684\u957f\u5ea6\u9884\u6d4b\u5668\uff08\u7528\u4e8e\u53d6\u4ee3\u67a2\u7ebd\u70b9\u5339\u914d\uff09\u3002\u4f5c\u4e3a\u4efb\u610f\u957f\u5ea6\u5e8f\u5217\u5efa\u6a21\u7684\u7b80\u5355\u66ff\u4ee3\u65b9\u6848\uff0c\u957f\u5ea6\u9884\u6d4b\u5668\u901a\u8fc7\u9884\u6d4b\u6240\u9700\u7684\u67a2\u7ebd\u70b9\u6570\u91cfk\uff0c\u5e76\u8f93\u51fa\u524dk\u4e2a\u70b9\u4f5c\u4e3a\u6700\u7ec8\u7ed3\u679c\uff0c\u6765\u5b9e\u73b0\u52a8\u6001\u70b9\u6570\u76ee\u5efa\u3002\u7b2c2\u884c\u548c\u7b2c4\u884c\u7684\u6bd4\u8f83\u9a8c\u8bc1\u4e86\u52a8\u6001\u67a2\u7ebd\u70b9\u5e8f\u5217\u5339\u914d\u7684\u6709\u6548\u6027\uff0c\u5c24\u5176\u662f\u5728\u590d\u6742\u5f62\u72b6\u7684\u9053\u8def\u8fb9\u754c\u4e0a\uff0cAP\u63d0\u9ad8\u4e868.9%\u3002\u7b2c3\u884c\u548c\u7b2c4\u884c\u7684\u6bd4\u8f83\u9a8c\u8bc1\u4e86\u63d0\u51fa\u7684\u70b9\u7ebf\u63a9\u7801\u6a21\u5757\u7684\u6709\u6548\u6027\uff0cmAP\u63d0\u9ad8\u4e864.9%\u3002<\/p>\n\n\n\n

        Discussion of the point-line relation modeling<\/strong><\/h3>\n\n\n\n
        \"\"<\/figure>\n\n\n\n

        \u5728\u88685\u4e2d\uff0c\u6211\u4eec\u6bd4\u8f83\u4e86\u5efa\u6a21\u70b9\u7ebf\u5173\u7cfb\u7684\u4e0d\u540c\u65b9\u6cd5\uff0c\u5305\u62ec\u6211\u4eec\u63d0\u51fa\u7684\u70b9\u7ebf\u63a9\u7801\u6a21\u5757\u548c\u73b0\u6709\u7684\u57fa\u4e8e\u5c42\u6b21\u67e5\u8be2\u7684\u65b9\u6cd5\u3002\u7b2c\u4e00\u884c\u4ee3\u8868\u6ca1\u6709\u70b9\u7ebf\u5148\u9a8c\u77e5\u8bc6\u7684 baseline\uff0c\u5373\u4f7f\u7528 mask-attention based transformer \u4f5c\u4e3a\u7ebf\u611f\u77e5\u70b9\u89e3\u7801\u5668\u3002\u7b2c\u4e00\u884c\u548c\u7b2c\u4e8c\u884c\u7684\u6bd4\u8f83\u8868\u660e\uff0c\u5c42\u6b21\u67e5\u8be2\u8bbe\u8ba1\u5bf9\u4e8e\u52a8\u6001\u5e8f\u5217\u5efa\u6a21\u662f\u65e0\u76ca\u7684\u3002\u539f\u56e0\u5982\u4e0b\u6240\u8ff0\uff1a\u6211\u4eec\u5728\u4e0a\u56fe\u4e2d\u5c55\u793a\u4e86\u5c42\u6b21\u67e5\u8be2\u548c\u70b9\u67e5\u8be2\u3002\u5728\u5c42\u6b21\u67e5\u8be2\u8bbe\u8ba1\u4e2d\uff0c\u5177\u6709\u76f8\u540c\u7d22\u5f15\u7684\u70b9\u67e5\u8be2\u5171\u4eab\u76f8\u540c\u7684\u70b9\u5d4c\u5165\u3002\u7136\u800c\uff0c\u7531\u4e8e\u6211\u4eec\u4f7f\u7528\u52a8\u6001\u6570\u91cf\u7684\u70b9\u6765\u5efa\u6a21\u5730\u56fe\u5143\u7d20\uff0c\u5728\u4e0d\u540c\u7ebf\u9884\u6d4b\u4e2d\u5177\u6709\u76f8\u540c\u7d22\u5f15\u7684\u70b9\u4e4b\u95f4\u51e0\u4e4e\u6ca1\u6709\u5173\u7cfb\u3002\u6839\u636e\u4e0d\u540c\u7684\u9884\u6d4b\u70b9\u7684\u5206\u5e03\uff0c\u88ab\u6311\u9009\u5f62\u6210\u6700\u7ec8\u67a2\u7ebd\u70b9\u5e8f\u5217\u7684\u70b9\u7d22\u5f15\u5404\u4e0d\u76f8\u540c\u3002\u7b80\u800c\u8a00\u4e4b\uff0c\u5c42\u6b21\u5d4c\u5165\u662f\u7d22\u5f15\u76f8\u5173\u7684\uff0c\u8fd9\u4e0e\u6211\u4eec\u7684\u7d22\u5f15\u65e0\u5173\u7684\u95ee\u9898\u5efa\u6a21\u76f8\u51b2\u7a81\uff0c\u800c\u4e14\u53ef\u80fd\u4f1a\u5f15\u5165\u5173\u4e8e\u7d22\u5f15\u4fe1\u606f\u7684\u566a\u97f3\u3002\u56e0\u6b64\uff0c\u4e3a\u4e86\u907f\u514d\u7d22\u5f15\u76f8\u5173\u7684\u4fe1\u606f\u5d4c\u5165\uff0c\u6211\u4eec\u91c7\u7528\u539f\u59cb\u7684\u70b9\u67e5\u8be2\uff0c\u5e76\u5c06\u4ece\u5c5e\u548c\u51e0\u4f55\u70b9\u7ebf\u5173\u7cfb\u5d4c\u5165\u5230\u6ce8\u610f\u529b\u63a9\u7801\u4e2d\u3002<\/p>\n\n\n\n

        Discussion of different AP thresholds.<\/strong><\/h3>\n\n\n\n
        \"\"<\/figure>\n\n\n\n

        \u4e0d\u540c\u7684 CD \u9608\u503c\u8bbe\u7f6e\u4ee3\u8868\u7740\u8bc4\u4f30\u534f\u8bae\u4e2d\u5bf9\u5730\u56fe\u7cbe\u5ea6\u7684\u4e0d\u540c\u5bb9\u5fcd\u7a0b\u5ea6\u3002\u8003\u8651\u5230\u5b9e\u9645\u81ea\u52a8\u9a7e\u9a76\u7684\u5e94\u7528\u4e2d\uff0c\u9ad8\u7cbe\u5ea6\u5730\u56fe\u901a\u5e38\u9700\u8981\u5398\u7c73\u7ea7\u7684\u4fe1\u606f\uff0c\u56e0\u6b64\u6211\u4eec\u8ba4\u4e3a\u5728\u66f4\u4e25\u683c\u7684\u9608\u503c\u4e0b\u7684\u6539\u8fdb\u66f4\u52a0\u5b9e\u9645\u3002\u4e0eHDMapNet \u76f8\u6bd4\uff0cVectorMapNet \u548c MapTR \u5728\u7b80\u5355\u9608\u503c\u4e0b\uff0c\u59821.0m\u30011.5m\uff0c\u6709\u663e\u8457\u6539\u8fdb\uff0c\u4f46\u5728\u66f4\u4e25\u683c\u7684\u573a\u666f\u4e0b\uff0c\u59820.2m\u30010.5m\uff0c\u5b83\u4eec\u7684\u6539\u8fdb\u5e45\u5ea6\u8fc5\u901f\u4e0b\u964d\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u65e0\u8bba\u91c7\u7528\u54ea\u79cd\u9608\u503c\uff0c PivotNet \u603b\u662f\u53d6\u5f97\u66f4\u597d\u7684\u7ed3\u679c\uff0c\u8fd9\u663e\u793a\u4e86\u6211\u4eec\u6846\u67b6\u7684\u9c81\u68d2\u6027\u3002<\/p>\n\n\n\n

        Visualization<\/strong><\/h3>\n\n\n\n
        \"\"<\/figure>\n\n\n\n
        \"\"<\/figure>\n\n\n\n

        \u6211\u4eec\u5728\u56fe5\u4e2d\u5c55\u793a\u4e86\u5728\u4e0d\u540c\u73af\u5883\u6761\u4ef6\u4e0b\u4e0eSOTA\u65b9\u6cd5\u7684\u5b9a\u6027\u6bd4\u8f83\u7ed3\u679c,\u4ece\u5de6\u81f3\u53f3\u5206\u522b\u4e3a\u591a\u89c6\u89d2\u56fe\u7247\u3001HDMapNet\u3001MapTR \u4e0e PivotNet \u7684\u6a21\u578b\u8f93\u51fa\u3002\u4e0e HDMapNet \u7684\u76f8\u6bd4\uff0cPivotNet \u9664\u4e86\u907f\u514d\u4e86\u590d\u6742\u7684\u540e\u5904\u7406\uff0c\u8fd9\u79cd\u57fa\u4e8e\u67a2\u7ebd\u70b9\u77e2\u91cf\u5316\u7684\u65b9\u6cd5\u6bd4\u57fa\u4e8e\u5206\u5272\u7684\u65b9\u6cd5\u66f4\u51c6\u786e\u5730\u9884\u6d4b\u4e86\u7aef\u70b9\u3002\u4e0e MapTR \u76f8\u6bd4\uff0c\u6211\u4eec\u7684\u65b9\u6cd5\u6bd4\u57fa\u4e8e\u5747\u5300\u70b9\u7684\u65b9\u6cd5\u80fd\u66f4\u5e73\u6ed1\u51c6\u786e\u5730\u8868\u8fbe\u5730\u56fe\u5f62\u72b6\uff0c\u4f8b\u5982\u76f4\u7ebf\u3001\u5706\u89d2\u548c\u76f4\u89d2\u3002\u6b64\u5916\uff0cPivotNet\u5728\u5efa\u6a21\u8fc7\u7a0b\u4e2d\u9700\u8981\u66f4\u5c11\u7684\u70b9\uff0c\u8868\u8fbe\u66f4\u52a0\u7d27\u51d1\u3002 \u4e0e GroundTruth \u7684\u6bd4\u8f83\u53ef\u4ee5\u770b\u51fa\uff0cPivotNet \u5bf9\u5404\u79cd\u9a7e\u9a76\u573a\u666f\u5177\u6709\u9c81\u68d2\u6027\uff0c\u5e76\u4e14\u80fd\u5728\u4e0d\u540c\u73af\u5883\u4e2d\u4fdd\u6301\u826f\u597d\u7684\u6027\u80fd\u3002\u5373\u4f7f\u5728\u591c\u665a\uff0c\u8f66\u8f86\u9644\u8fd1\u7684\u5730\u56fe\u4e5f\u4e0e\u5b9e\u9645\u60c5\u51b5\u975e\u5e38\u63a5\u8fd1\u3002<\/p>\n\n\n\n

        \u7ed3\u8bba\uff1a<\/strong><\/h2>\n\n\n\n

        \u672c\u6587\u4e3b\u8981\u5173\u6ce8\u5c40\u90e8\u9ad8\u7cbe\u5730\u56fe\u6784\u5efa\u4efb\u52a1\uff0c\u5e76\u5c06\u8be5\u4efb\u52a1\u5efa\u6a21\u4e3a\u4e00\u4e2a\u70b9\u96c6\u9884\u6d4b\u95ee\u9898\u3002\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u4e2a\u7aef\u5230\u7aef\u7684\u6846\u67b6\uff0c\u5b83\u4f7f\u7528\u7edf\u4e00\u7684\u57fa\u4e8e\u67a2\u7ebd\u70b9\u7684\u77e2\u91cf\u5316\u8868\u8fbe\u6765\u6784\u5efa\u7cbe\u51c6\u4e14\u7d27\u51d1\u7684\u9ad8\u6e05\u5730\u56fe\uff0c\u540c\u65f6\u5728\u7f51\u7edc\u5efa\u6a21\u4e2d\u5d4c\u5165\u4e86\u70b9\u7ebf\u51e0\u4f55\u548c\u4ece\u5c5e\u5173\u7cfb\u7684\u5148\u9a8c\u3002\u901a\u8fc7\u5f15\u5165\u70b9\u7ebf\u63a9\u7801\u6a21\u5757\u3001\u67a2\u7ebd\u70b9\u52a8\u6001\u5339\u914d\u6a21\u5757\u548c\u52a8\u6001\u77e2\u91cf\u5316\u5e8f\u5217\u635f\u5931\uff0c\u6211\u4eec\u63d0\u51fa\u7684 PivotNet \u8fbe\u5230\u4e86 SOTA \u7684\u7ed3\u679c\uff0c\u5e76\u4e3a\u672a\u6765\u7684\u7814\u7a76\u63d0\u4f9b\u4e86\u65b0\u7684\u89c6\u89d2\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

        \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/ […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3,1],"tags":[],"class_list":["post-7128","post","type-post","status-publish","format-standard","hentry","category-technology-frontier","category-home"],"views":1,"_links":{"self":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/7128"}],"collection":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/comments?post=7128"}],"version-history":[{"count":1,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/7128\/revisions"}],"predecessor-version":[{"id":7144,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/7128\/revisions\/7144"}],"wp:attachment":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/media?parent=7128"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/categories?post=7128"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/tags?post=7128"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}