{"id":7280,"date":"2024-01-03T16:06:15","date_gmt":"2024-01-03T08:06:15","guid":{"rendered":"http:\/\/192.168.10.115\/?p=7280"},"modified":"2024-01-03T16:07:04","modified_gmt":"2024-01-03T08:07:04","slug":"2024-01-03","status":"publish","type":"post","link":"http:\/\/222.128.65.89:18086\/index.php\/2024\/01\/03\/7280\/","title":{"rendered":"2024-01-03 \u6ca1\u6709\u9ad8\u7cbe\u5730\u56fe\u5982\u4f55\u5b9e\u73b0L4\u7ea7\u522b\u4ee5\u4e0a\u81ea\u52a8\u9a7e\u9a76"},"content":{"rendered":"\n

\u4f17\u6240\u5468\u77e5,\u7531\u4e8e\u79bb\u7ebf\u9ad8\u7cbe\u5730\u56feHigh-Definition Map(HDMap)\u5236\u4f5c\u6210\u672c\u9ad8\u6602,\u4e14\u96be\u4ee5\u5b9e\u65f6\u66f4\u65b0,\u57fa\u4e8e\u81ea\u8f66\u5b9e\u65f6\u611f\u77e5\u751f\u6210\u7684\u5728\u7ebf\u5c40\u90e8\u5730\u56fe\u5df2\u7ecf\u9010\u6e10\u6210\u4e3a\u66ff\u4ee3\u89e3\u51b3\u65b9\u6848\u6216\u662f\u5197\u4f59\u65b9\u6848,\u56e0\u4e3a\u79bb\u7ebf\u5730\u56fe<\/a>\u7cbe\u5ea6\u8f83\u9ad8\u4e14\u5177\u5907\u5168\u5c40\u4fe1\u606f,\u5728\u7ebf\u5730\u56fe<\/a><\/p>\n\n\n\n

\u867d\u7136\u4e0d\u5177\u5907\u5168\u5c40\u4fe1\u606f\u4f46\u5177\u6709\u5b9e\u65f6\u6027,\u4e8c\u8005\u53ef\u4ee5\u4e92\u4e3a\u8865\u5145\u3002<\/p>\n\n\n\n

\u535a\u4e3b\u5728\u8fd9\u4e2a\u9886\u57df\u505a\u4e86\u7b80\u8981\u7684\u8c03\u7814,\u603b\u7ed3\u6bd4\u8f83\u7ecf\u5178\u7684\u65b9\u6848\u4e3b\u8981\u6709\u4e09\u4e2a:\u6e05\u534e\u8d75\u884cMARS Lab(\u548c\u8fd9\u4e2a\u7ec4\u6760\u4e0a\u4e86)\u7684HDMapNet, VectorMap,\u548c\u534e\u4e2d\u79d1\u6280\u5927\u5b66<\/a>\u4e0e\u5730\u5e73\u7ebf\u5408\u4f5c\u7684MapTR,\u89c9\u5f97\u4e09\u4e2a\u6a21\u578b\u5728\u6846\u67b6\u548c\u601d\u8def\u4e0a\u4e5f\u6709\u5185\u5728\u8054\u7cfb\u548c\u6f14\u8fdb\u7684\u8fc7\u7a0b,\u5e76\u4e14\u548cBEV\u52a8\u6001\u76ee\u6807\u68c0\u6d4b<\/a><\/p>\n\n\n\n

\u4e5f\u6709\u76f8\u901a\u4e4b\u5904,\u5728\u6b64\u505a\u4e2a\u7b80\u8981\u7684\u5bf9\u6bd4\u548c\u5206\u6790,\u5e76\u52a0\u5165\u81ea\u5df1\u7684\u601d\u8003\u3002<\/p>\n\n\n\n

\u9996\u5148,\u4f20\u7edf\u5728\u7ebf\u5c40\u90e8\u5730\u56fe\u751f\u6210\u65b9\u5f0f\u5927\u591a\u57fa\u4e8e\u591a\u4e2a\u900f\u89c6\u89c6\u89d2\u7684\u8bed\u4e49\u5206\u5272,\u901a\u8fc7IPM\u6295\u5f71\u548c\u62fc\u63a5\u7b49\u65b9\u5f0f\u8f6c\u6362\u4e3a\u9e1f\u77b0\u89c6\u89d2,\u518d\u7ecf\u8fc7\u6bd4\u8f83\u590d\u6742\u7684\u540e\u5904\u7406\u5f97\u5230\u6700\u7ec8\u9700\u8981\u7684\u5730\u56fe\u8981\u7d20\u548c\u62d3\u6251\u7ed3\u6784,\u8fd9\u91cc\u4e0d\u505a\u8fc7\u591a\u4ecb\u7ecd.BEV\u8303\u5f0f\u51fa\u73b0\u540e,\u53ef\u4ee5\u5148\u901a\u8fc7\u591a\u79cd\u65b9\u6cd5\u5f97\u5230BEV feature,\u7136\u540e\u5728BEV feature\u4e0a\u8fdb\u884c\u540e\u7eed\u64cd\u4f5c,\u4f7f\u6574\u4e2a\u8fc7\u7a0b\u7b80\u6d01\u4e86\u5f88\u591a.HDMapNet \u662f\u76f8\u5bf9\u8f83\u65e9\u7684BEV\u5728\u7ebf\u5c40\u90e8\u5730\u56fe\u6784\u5efa\u65b9\u6848\u3002<\/p>\n\n\n\n

\n

\u5173\u6ce8\u77e5\u4e4e@\u81ea\u52a8\u9a7e\u9a76\u4e4b\u5fc3<\/a>\uff0c\u7b2c\u4e00\u65f6\u95f4\u83b7\u53d6\u81ea\u52a8\u9a7e\u9a76\u611f\u77e5\/\u5b9a\u4f4d\/\u878d\u5408\/\u89c4\u63a7\u7b49\u884c\u4e1a\u6700\u65b0\u5185\u5bb9<\/strong><\/p>\n<\/blockquote>\n\n\n\n

1.HDMapNet (2022.3)[1]<\/strong><\/h2>\n\n\n\n

[1] HDMapNet: An Online HD Map Construction and Evaluation Framework<\/p>\n\n\n\n

\"\"
HDMapNet\u6846\u67b6<\/figcaption><\/figure>\n\n\n\n

HDMapNet\u5ef6\u7eed\u4e86\u4f20\u7edf\u65b9\u6cd5\u4e0a\u57fa\u4e8e\u8bed\u4e49\u5206\u5272\u7684\u601d\u8def.\u9996\u5148\u4f7f\u7528VPN[2]\u65b9\u6cd5\u751f\u6210BEV Feature,\u7136\u540e\u5730\u56fehead\u5206\u4e3a\u4e09\u4e2a\u5206\u652f,\u7b2c\u4e00\u4e2a\u7ed9\u51fa\u5728bev feature\u4e0a\u505a\u8bed\u4e49\u5206\u5272\u7684\u7ed3\u679c,\u5c06\u5173\u6ce8\u7684\u8f66\u9053\u7ebf,\u9053\u8def\u8fb9\u7ebf,\u8def\u53e3\u7b49\u4f5c\u51fa\u5206\u5272,\u8fd9\u4e2a\u6bd4\u8f83\u597d\u7406\u89e3,\u4f46\u662f\u53ea\u6709\u5206\u5272\u7ed3\u679c\u80af\u5b9a\u662f\u4e0d\u591f\u7684,\u6211\u4eec\u77e5\u9053\u5728\u9ad8\u7cbe\u5730\u56fe\u7684\u8868\u8fbe\u5f62\u5f0f\u4e2d,\u8f66\u9053\u7ebf\u7b49\u5730\u56fe\u8981\u7d20\u662f\u7528\u5b9e\u4f8b(Instance)\u6765\u8868\u793a\u7684,\u5177\u4f53\u5728\u8f66\u9053\u7ebf\u4e0a\u4ee3\u8868\u4e00\u6bb5\u8f66\u9053\u7ebf,\u5185\u90e8\u7531\u7ebf\u6bb5\u4e0a\u5177\u6709\u4e00\u5b9a\u987a\u5e8f\u7684\u70b9\u6765\u8868\u793a,\u5b9e\u4f8b\u4e4b\u95f4\u4e5f\u5b58\u5728\u4e00\u5b9a\u7684\u62d3\u6251\u5173\u7cfb,\u6240\u4ee5\u6211\u4eec\u9700\u8981\u7684\u662f\u4e00\u4e2a\u7531\u5b9e\u4f8b\u6784\u6210\u7684\u9002\u91cf\u56fe.\u8fd9\u91cc\u4f5c\u8005\u901a\u8fc7\u540e\u4e24\u4e2a\u5206\u652f\u548c\u540e\u5904\u7406\u6765\u5f97\u5230\u9700\u8981\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n

[2] https:\/\/arxiv.org\/abs\/1906.03560<\/a><\/p>\n\n\n\n

\u7b2c\u4e8c\u4e2a\u5206\u652f\u7ed9\u51fa\u5b9e\u4f8b\u5d4c\u5165(Instance embedding<\/a>),\u8bba\u6587\u91cc\u9762\u4ecb\u7ecd\u6bd4\u8f83\u8be6\u7ec6,\u6211\u7684\u7406\u89e3\u662f\u4e3aBEV Feature\u7684\u6bcf\u4e2a\u50cf\u7d20\u5206\u914d\u4e00\u4e2aembedding,\u7c7b\u4f3c\u4e8eNLP\u4e2d\u7684\u8bcd\u5411\u91cf,\u901a\u8fc7\u635f\u5931\u51fd\u6570<\/a><\/p>\n\n\n\n

\u7684\u8bbe\u8ba1,\u6700\u7ec8\u8bad\u7ec3\u51fa\u5c5e\u4e8e\u540c\u4e00\u5b9e\u4f8b\u76f8\u4f3c\u5ea6\u8f83\u9ad8,\u4e0d\u5c5e\u4e8e\u540c\u4e00\u52bf\u529b\u76f8\u4f3c\u5ea6\u8f83\u4f4e\u7684embedding,\u540e\u7eed\u901a\u8fc7\u5bf9embedding\u8fdb\u884c\u76f8\u4f3c\u6027\u5ea6\u91cf,\u5373\u53ef\u533a\u5206\u51fa\u4e0d\u540c\u7684\u5b9e\u4f8b.\u53c2\u8003\u4e0a\u56fe\u5bf9instance embedding\u7684\u53ef\u89c6\u5316. \u7b2c\u4e09\u4e2a\u5206\u652f\u4f1a\u7ed9\u51fa\u8f66\u9053\u7ebf\u4e0a\u6bcf\u4e2a\u70b9\u7684\u65b9\u5411,\u8fd9\u4e2a\u65b9\u5411\u662f\u7531\u76f8\u90bb\u4e24\u4e2a\u70b9\u7684\u8fde\u7ebf\u65b9\u5411\u786e\u5b9a\u7684,\u5e76\u4e14\u53ef\u4ee5\u6709\u4e24\u79cd\u53ef\u80fd\u6027,\u6a21\u578b\u90fd\u4f1a\u7ed9\u51fa,\u5e76\u4e14\u53ea\u6709\u771f\u6b63\u843d\u5728\u8f66\u9053\u7ebf\u4e2d\u5fc3\u7684\u70b9\u624d\u7ed9\u51fa\u65b9\u5411,\u8fd9\u4e00\u70b9\u901a\u8fc7\u5728\u8bad\u7ec3\u9636\u6bb5\u5c06\u4e0d\u843d\u5728\u8f66\u9053\u7ebf\u4e2d\u5fc3\u7684\u70b9\u7684\u65b9\u5411\u771f\u503c\u8d4b\u4e3a0\u6765\u5b9e\u73b0\u3002<\/p>\n\n\n\n

\u6709\u4e86\u8bed\u4e49\u5206\u5272<\/a>,\u5b9e\u4f8b\u5d4c\u5165\u548c\u65b9\u5411,\u5373\u53ef\u5728\u540e\u5904\u7406\u9636\u6bb5\u5efa\u7acb\u4e09\u8005\u7684\u8054\u7cfb,\u6700\u7ec8\u7ed9\u51fa\u77e2\u91cf\u5730\u56fe.\u53ef\u4ee5\u770b\u51fa\u8fd9\u79cd\u65b9\u6cd5\u662f\u6bd4\u8f83\u76f4\u89c2,\u5bb9\u6613\u7406\u89e3\u7684,\u4f46\u52a3\u52bf\u4e5f\u662f\u663e\u800c\u6613\u89c1\u7684,\u8ba1\u7b97\u91cf\u8f83\u5927,\u8017\u65f6\u8f83\u957f,\u4e14\u4e0d\u591fEnd-to-End,\u90e8\u7f72\u96be\u5ea6\u9ad8,\u7f3a\u5c11\u4e86BEV\u8303\u5f0f\u7684\u4f18\u7f8e\u6c14\u8d28,\u4f46\u4f5c\u4e3a\u57fa\u7ebf\u4e5f\u662f\u5f88\u96be\u5f97\u4e86\u3002<\/p>\n\n\n\n

\u90a3\u80fd\u4e0d\u80fd\u901a\u8fc7\u6a21\u578b\u76f4\u63a5\u5f97\u5230\u77e2\u91cf\u5730\u56fe<\/a>\u800c\u4e0d\u9700\u8981\u540e\u5904\u7406\u5462?(\u611f\u89c9\u5404\u4e2a\u65b9\u5411\u7684\u6f14\u8fdb\u90fd\u662f\u8fd9\u4e2a\u5957\u8def)\u8d75\u884c<\/a>\u5927\u4f6c\u6700\u540e\u7ed9\u51fa\u4e86VectorMapNet\u3002<\/p>\n\n\n\n

2.VectorMapNet(2022)[3]<\/strong><\/h2>\n\n\n\n

[3] VectorMapNet: End-to-end Vectorized HD Map Learning<\/p>\n\n\n\n

\"\"
VectorMapNet<\/figcaption><\/figure>\n\n\n\n

VectorMapNet\u6452\u5f03\u4e86\u901a\u8fc7\u8bed\u4e49\u5206\u5272\u5f97\u5230\u5730\u56fe\u7684\u601d\u8def,\u800c\u662f\u4f7f\u7528\u76ee\u6807\u68c0\u6d4b\u7684\u601d\u8def(\u4e0d\u77e5\u9053\u662f\u4e0d\u662f\u7b2c\u4e00\u4e2a\u8fd9\u6837\u505a\u7684),\u5c06\u4e0d\u89c4\u5219\u7684\u5730\u56fe\u8981\u7d20\u5b9e\u4f8b\u62bd\u8c61\u6210\u51e0\u4e2a\u5173\u952e\u70b9(\u53ef\u4ee5\u662fbounding box,\u6700\u5927\u6700\u5c0f\u503c,\u8d77\u70b9\u4e2d\u70b9\u7ec8\u70b9\u7b49\u7b49),\u8fd9\u6837\u5c31\u548c\u76ee\u6807\u68c0\u6d4b\u7684\u8303\u5f0f\u5bf9\u9f50\u4e86!\u662f\u4e0d\u662f\u5f88\u60ca\u8273.\u5f53\u7136\u53ea\u6709\u5173\u952e\u70b9\u80af\u5b9a\u662f\u4e0d\u591f\u7684,\u8bba\u6587\u4e2d\u540e\u7eed\u901a\u8fc7Polyline Generator,\u5373\u751f\u6210\u7684\u65b9\u5f0f\u9010\u4e2a\u5f97\u5230\u6700\u7ec8\u9700\u8981\u7684\u6709\u5e8f\u70b9\u96c6,\u4e0d\u9700\u8981\u540e\u5904\u7406,\u5b9e\u73b0\u4e86\u7aef\u5230\u7aef\u751f\u6210\u77e2\u91cf\u5730\u56fe\u3002<\/p>\n\n\n\n

\"\"
\u5730\u56fe\u8981\u7d20\u7684\u5173\u952e\u70b9<\/figcaption><\/figure>\n\n\n\n

\u5177\u4f53\u6765\u8bb2,VectorMapNet\u4e3b\u8981\u5206\u4e3a\u4e09\u4e2a\u9636\u6bb5,\u7b2c\u4e00\u9636\u6bb5\u662f\u901a\u8fc7\u591a\u9ad8\u5ea6(\u22121m, 0m, 1m, 2m)\u7684IPM\u65b9\u5f0f\u5bf9\u5730\u56fe\u8981\u7d20\u6295\u5f71\u5230BEV\u7a7a\u95f4,\u8fd9\u91cc\u5e94\u8be5\u662f\u5229\u7528\u5730\u56fe\u8981\u7d20\u9759\u6001\u4e14\u7a33\u5b9a\u7684\u7279\u70b9,\u6ca1\u6709\u4f7f\u7528\u73b0\u5728\u6d41\u884c\u7684\u57fa\u4e8e\u5b66\u4e60\u7684\u65b9\u5f0f.\u7b2c\u4e8c\u9636\u6bb5\u5373\u5730\u56fe\u8981\u7d20\u68c0\u6d4b\u9636\u6bb5,\u4f7f\u7528\u4e86\u7ecf\u5178\u7684DETR[4]\u548cDeformable DETR[5]\u8303\u5f0f,\u5f97\u5230\u5730\u56fe\u8981\u7d20\u7684\u5173\u952e\u70b9,\u53ea\u662f\u8fd9\u91ccquery\u7684\u8bbe\u8ba1\u6bd4\u8f83\u7279\u6b8a.\u5728\u52a8\u6001\u76ee\u6807\u68c0\u6d4b\u4e2d,\u6211\u4eec\u4e00\u822c\u53ea\u9700\u8981\u5b9a\u4e49\u51e0\u767e\u4e2a\u5730\u4f4d\u76f8\u540c\u7684query,\u76f4\u63a5\u8f93\u5165transformer<\/a><\/p>\n\n\n\n

\u5c31\u53ef\u4ee5,\u4f46\u5730\u56fe\u8981\u7d20\u548c\u52a8\u6001\u76ee\u6807\u6709\u70b9\u4e0d\u540c,\u6211\u4eec\u9996\u5148\u8981\u7ed9\u51fa\u4e00\u5b9a\u6570\u91cf(\u6570\u91cf\u4e00\u822c\u4e0d\u4f1a\u5f88\u591a)\u7684\u5b9e\u4f8b,\u5047\u8bbe\u6700\u591am\u4e2a,\u6bcf\u4e2a\u5b9e\u4f8b\u53c8\u5305\u62ec\u4e00\u5b9a\u6570\u91cf\u7684\u5173\u952e\u70b9,\u5047\u8bbe\u56fa\u5b9an\u4e2a,\u6700\u7ec8\u9700\u8981\u5f97\u5230m*n\u4e2a\u70b9,\u4f46\u8fd9\u4e9b\u70b9\u4e0d\u662f\u76f8\u4e92\u72ec\u7acb\u7684,\u800c\u662f\u6709\u4e00\u4e2a\u4e8c\u7ea7\u7684\u7ee7\u627f\u5173\u7cfb.\u5982\u679c\u6211\u4eec\u8bbe\u5b9a\u6bcf\u4e2aquery\u9884\u6d4b\u4e00\u4e2a\u70b9,\u4e00\u5171\u9700\u8981\u5b9a\u4e49m*n\u4e2aquery, \u5982\u679c\u76f4\u63a5\u628am*n\u4e2aquery\u8f93\u5165transformer,\u7ed3\u679c\u80af\u5b9a\u65e0\u6cd5\u4f53\u73b0\u8fd9\u79cd\u7ee7\u627f\u5173\u7cfb. \u6587\u4e2d\u7ed9\u51fa\u4e00\u4e2a\u6bd4\u8f83\u5de7\u5999\u7684\u65b9\u6cd5,\u5373\u5148\u5b9a\u4e49m\u4e2a\u5b9e\u4f8bquery,\u5047\u8bbe\u7ef4\u5ea6\u4e3adim, \u5927\u5c0f\u4e3am * dim, \u518d\u5b9a\u4e49n\u4e2a\u5173\u952e\u70b9query, \u5927\u5c0f\u4e3an* dim,\u4f7f\u7528\u5e7f\u64ad\u7684\u65b9\u5f0f\u8ba9\u4e8c\u8005\u76f8\u52a0,\u5373\u53ef\u5f97\u5230m*n\u4e2a\u540c\u65f6\u62e5\u6709\u5b9e\u4f8b\u4fe1\u606f\u548c\u5173\u952e\u70b9\u4fe1\u606f\u7684query,\u4f9d\u6b21\u8f93\u5165transformer.\u7531\u4e8e\u5c5e\u4e8e\u540c\u4e00\u4e2ainstance\u7684query\u5b9e\u9645\u4e0a\u4ee3\u8868\u4e00\u4e2a\u6574\u4f53,\u53ea\u80fd\u6709\u4e00\u79cd\u7c7b\u522b,\u6240\u4ee5\u7c7b\u522b\u5206\u652f\u8f93\u5165\u7684\u662f\u540c\u4e00\u4e2ainstance\u7684query\u7684\u5e73\u5747\u503c,\u5f97\u5230\u5206\u7c7bscore.\u6700\u7ec8\u7b5b\u9009\u51fascore\u9ad8\u4e8e\u9608\u503c\u7684\u591a\u7ec4query,\u5373\u53ef\u5f97\u5230\u6211\u4eec\u9700\u8981\u7684\u4e8c\u7ea7\u7ed3\u679c\u3002<\/p>\n\n\n\n

[4] End-to-End Object Detection with Transformers<\/p>\n\n\n\n

[5] Deformable DETR: Deformable Transformers for End-to-End Object Detection<\/p>\n\n\n\n

\u8bf4\u5230\u8fd9\u91cc,\u719f\u6089DETR\u7684\u540c\u5b66\u80af\u5b9a\u4f1a\u60f3\u5230,\u4e3a\u4ec0\u4e48\u8981\u8fd9\u4e48\u9ebb\u70e6\u5462,\u6211\u4eec\u53ea\u5b9a\u4e49m\u4e2a\u5b9e\u4f8bquery,\u8ba9\u6bcf\u4e2aquery\u9884\u6d4bn\u4e2a\u5173\u952e\u70b9\u4e0d\u5c31\u53ef\u4ee5\u5417,\u4e3a\u4ec0\u4e48\u8981\u4e00\u4e2aquery\u9884\u6d4b\u4e00\u4e2a\u5173\u952e\u70b9,\u5b9a\u4e49\u4e24\u7ea7\u7684query\u5462,\u6211\u7684\u7b2c\u4e00\u53cd\u5e94\u4e5f\u662f\u8fd9\u4e2a,\u4f46\u4ed4\u7ec6\u60f3\u4e00\u4e0b,\u5730\u56fe\u8981\u7d20\u548c\u52a8\u6001\u76ee\u6807\u7684\u533a\u522b\u8fd8\u662f\u5f88\u660e\u663e\u7684.\u9996\u5148\u4e00\u5e27\u56fe\u7247\u4e2d\u7684\u5730\u56fe\u5143\u7d20\u5b9e\u4f8b\u4e0d\u53ef\u80fd\u5f88\u591a\u7684,\u4e00\u822c\u53ea\u6709\u51e0\u4e2a\u6216\u66f4\u591a\u4e00\u70b9,\u6211\u4eec\u6ca1\u6709\u5fc5\u8981\u5b9a\u4e49\u8fc7\u591a\u7684query,\u591a\u4e86\u53cd\u800c\u53ef\u80fd\u5f71\u54cd\u6548\u679c,\u8fd9\u6837\u9650\u5236\u4e86query\u7684\u6570\u91cf.\u5176\u6b21,\u6bcf\u4e2a\u5b9e\u4f8b\u7531\u4e8e\u957f\u5ea6\u6bd4\u8f83\u957f,\u4e14\u5f62\u72b6\u4e0d\u89c4\u5219,\u5185\u90e8\u4fe1\u606f\u5176\u5b9e\u662f\u6bd4\u8f83\u591a\u7684,\u8fd9\u6837\u6bcf\u4e2a\u5173\u952e\u70b9\u76f8\u5bf9\u72ec\u7acb,\u4e0d\u50cf\u52a8\u6001\u76ee\u6807\u5f62\u72b6\u6bd4\u8f83\u89c4\u5219,\u6bcf\u4e2a\u5173\u952e\u70b9\u5177\u6709\u5f3a\u76f8\u5173\u6027,\u53ea\u9700\u8981\u5b9a\u4e49\u591a\u4e2aquery\u5373\u53ef\u8986\u76d6\u591a\u79cd\u76ee\u6807\u7684\u4fe1\u606f.\u6240\u4ee5\u5982\u679c\u5730\u56fe\u5143\u7d20\u68c0\u6d4b\u53ea\u5b9a\u4e49\u6bd4\u8f83\u5c11\u7684m\u4e2a\u5b9e\u4f8bquery\u4e0d\u662f\u4e0d\u53ef\u4ee5,\u6548\u679c\u80af\u5b9a\u4f1a\u5dee\u4e00\u70b9,\u4f5c\u8005\u5e94\u8be5\u4e5f\u6709\u8fd9\u6837\u7684\u8003\u8651.\u800c\u4e14\u5982\u679c\u6bcf\u4e2a\u5b9e\u4f8b\u4e0d\u662f\u5bf9\u5e94\u5c11\u91cf\u5173\u952e\u70b9\u800c\u662f\u591a\u4e2a\u70b9,\u4e24\u79cd\u65b9\u6cd5\u7684\u533a\u522b\u4f1a\u66f4\u52a0\u660e\u663e\u3002<\/p>\n\n\n\n

\u7b2c\u4e09\u9636\u6bb5\u662fPolyline\u751f\u6210\u9636\u6bb5,\u5173\u4e8e\u751f\u6210(Generator)\u7684\u9886\u57df,\u4e5f\u5c31\u662f\u73b0\u5728\u5927\u706b\u7684GPT-4<\/a>\u7684\u57fa\u7840\u6280\u672f\u535a\u4e3b\u4e86\u89e3\u4e0d\u662f\u5f88\u591a,\u7b80\u5355\u5730\u8bf4\u5c31\u662f\u901a\u8fc7\u81ea\u56de\u5f52<\/a>\u7684\u65b9\u5f0f,\u4ece\u7b2c\u4e8c\u9636\u6bb5\u7684\u5173\u952e\u70b9\u751f\u6210\u5b8c\u6574\u7684\u5730\u56fe\u8981\u7d20\u6709\u5e8f\u70b9\u96c6.\u8fd9\u91cc\u70b9\u96c6\u4e2d\u70b9\u7684\u6570\u91cf\u662f\u4e0d\u56fa\u5b9a\u7684,\u901a\u8fc7\u751f\u6210End of Sequence token (EOS)\u6807\u5fd7\u4f4d\u4ee3\u8868\u7ed3\u675f,\u4e5f\u662f\u7c7b\u4f3c\u4e8eNLP\u7684\u6280\u672f,\u53ef\u89c1\u76ee\u524dNLP\u548ccv\u7684\u7ed3\u5408\u8d8a\u6765\u8d8a\u7d27\u5bc6,\u56e0\u4e3a\u5185\u6838\u90fd\u7528\u4e86\u540c\u6837\u7684transformer\u5927\u6740\u5668.\u6570\u91cf\u4e0d\u56fa\u5b9a\u662f\u751f\u6210\u65b9\u6cd5\u7684\u4f18\u52bf,\u53ef\u4ee5\u8868\u5f81\u4e0d\u540c\u957f\u5ea6\u7684\u5730\u56fe\u8981\u7d20,\u4f46\u5b83\u7684\u52a3\u52bf\u5c31\u662f\u751f\u6210\u8fc7\u7a0b\u662f\u5e8f\u5217\u6027\u7684,\u53ea\u80fd\u505a\u5230\u591a\u4e2ainstance\u7684\u5e76\u884c,instance\u5185\u90e8\u65e0\u6cd5\u5e76\u884c,\u964d\u4f4e\u4e86\u6548\u7387.\u8fd8\u6709\u4e00\u4e2a\u9690\u85cf\u7684\u52a3\u52bf\u5c31\u662f\u6709\u5e8f\u6027,\u5176\u5b9e\u5730\u56fe\u8981\u7d20\u7684\u70b9\u867d\u7136\u662f\u6709\u5e8f\u7684,\u4f46\u8d77\u70b9\u4e0d\u662f\u552f\u4e00\u7684,\u4f8b\u5982\u8f66\u9053\u7ebf\u53ef\u4ee5\u6709\u4e24\u79cd\u8d77\u70b9,\u95ed\u5408\u7684\u5143\u7d20\u5982\u8def\u53e3,\u4efb\u610f\u4e00\u4e2a\u70b9\u90fd\u53ef\u4ee5\u4f5c\u4e3a\u8d77\u70b9,\u800c\u751f\u6210\u7684\u65b9\u5f0f\u53ea\u80fd\u6709\u552f\u4e00\u7684\u8d77\u70b9,\u76f8\u5f53\u4e8e\u5bf9\u6b64\u505a\u4e86\u4e0d\u5fc5\u8981\u4e14\u4e0d\u5408\u7406\u7684\u7ea6\u675f,\u5bf9\u7ed3\u679c\u80af\u5b9a\u6709\u4e00\u5b9a\u7684\u5f71\u54cd\u3002\u6709\u6ca1\u6709\u66f4\u52a0\u7b80\u6d01\u800c\u5408\u7406\u7684\u65b9\u6848\u5462?\u90a3\u5c31\u662f\u4e0b\u9762\u7684MapTR\u4e86\u3002<\/p>\n\n\n\n

3.MapTR(2023.1)[6]<\/strong><\/h2>\n\n\n\n

[6] MapTR: Structured Modeling and Learning for Online Vectorized HD Map Construction<\/p>\n\n\n\n

\"\"
MapTR<\/figcaption><\/figure>\n\n\n\n

MapTR\u53ef\u4ee5\u8bf4\u548cVectorMapNet\u8054\u7cfb\u76f8\u5f53\u7d27\u5bc6\u4e86,\u4ece\u5927\u7684\u6846\u67b6\u6765\u770b,\u672c\u8d28\u4e0a\u5176\u5b9e\u5c31\u662f\u53ea\u4f7f\u7528\u4e86VectorMapNet\u7684\u5730\u56fe\u8981\u7d20\u68c0\u6d4b\u6a21\u5757,\u800c\u4e22\u5f03\u4e86\u540e\u9762\u7684\u751f\u6210\u6a21\u5757.\u4e5f\u5c31\u662f\u8bf4,MapTR\u6ca1\u6709\u5148\u751f\u6210\u5730\u56fe\u8981\u7d20\u5173\u952e\u70b9,\u800c\u662f\u76f4\u63a5\u4f7f\u7528\u57fa\u4e8eDETR\u548cDeformable DETR\u7684\u76ee\u6807\u68c0\u6d4b\u65b9\u5f0f\u751f\u6210\u6700\u7ec8\u7684\u5730\u56fe\u8981\u7d20\u70b9,\u4f46\u662f\u662f\u56fa\u5b9a\u957f\u5ea6\u7684,\u6587\u4e2d\u662f20\u4e2a\u70b9.query\u7684\u8bbe\u8ba1\u65b9\u5f0f\u4e0eVectorMapNet\u5b8c\u5168\u4e00\u81f4.\u6839\u636e\u4e0a\u6587\u7684\u5206\u6790,\u5982\u679c\u53ea\u8bbe\u8ba1m\u4e2a\u5b9e\u4f8bquery,\u90a3\u4e00\u4e2aquery\u9700\u8981\u9884\u6d4b20\u4e2a\u6beb\u65e0\u89c4\u5219\u7684\u70b9,\u663e\u7136\u4fe1\u606f\u91cf\u662f\u4e0d\u591f\u7684,\u6240\u4ee5\u4e8c\u7ea7query<\/a>\u7684\u65b9\u5f0f\u66f4\u52a0\u663e\u5f97\u6709\u5fc5\u8981\u3002<\/p>\n\n\n\n

\u5f53\u7136MapTR\u76f8\u6bd4\u4e8eVectorMapNet\u8fd8\u6709\u5176\u4ed6\u7684\u6539\u8fdb.\u7b2c\u4e00,\u5b83\u662fpermutation-equivalent,\u7c7b\u4f3c\u4e8e\u5e73\u79fb\u4e0d\u53d8\u6027.\u8fd9\u662f\u76f8\u5bf9\u4e8eVectorMapNet\u7528\u751f\u6210\u7684\u65b9\u5f0f\u53ea\u80fd\u5f97\u5230\u552f\u4e00\u7684\u987a\u5e8f,MapTR\u53bb\u6389\u4e86\u8fd9\u4e2a\u4e0d\u5408\u7406\u7684\u7ea6\u675f.\u5177\u4f53\u65b9\u6848\u662f\u5c06\u5730\u56fe\u8981\u7d20\u5206\u4e3a\u4e24\u7c7b\u5904\u7406,\u4e00\u7c7b\u662fPolyline,\u5373\u5f00\u653e\u7684\u8981\u7d20,\u4e00\u7c7b\u662fPolygon,\u5373\u5c01\u95ed\u7684\u8981\u7d20,\u4e8c\u8005\u901a\u8fc7\u8d77\u70b9\u7ec8\u70b9\u7684\u8ddd\u79bb\u6765\u533a\u5206.\u5bf9\u4e8ePolyline,\u652f\u6301\u4e24\u79cd\u65b9\u5411,\u56e0\u4e3a\u8d77\u70b9\u6709\u4e24\u79cd\u53ef\u80fd,\u5bf9\u4e8ePolygon,\u652f\u63012k\u79cd\u65b9\u5411(k\u4ee3\u8868\u70b9\u96c6\u7684\u5927\u5c0f),\u56e0\u4e3a\u8d77\u70b9\u6709k\u79cd\u53ef\u80fd.\u5c06Polyline\u548cPolygon\u7684gt\u90fd\u8bbe\u8ba1\u62102k\u5927\u5c0f(Polyline\u8fdb\u884cpadding),\u4f7f\u9884\u6d4b\u503c\u4e0e\u6bcf\u4e00\u79cdgt\u90fd\u8fdb\u884c\u5339\u914d,\u53d6\u8bef\u5dee\u6700\u5c0f\u7684\u4e00\u79cd\u4f5c\u4e3a\u5339\u914d\u7ed3\u679c,\u8fd9\u5c31\u53bb\u6389\u4e86\u987a\u5e8f\u552f\u4e00\u7684\u7ea6\u675f.\u5b9e\u9a8c\u8868\u660epermutation-equivalent\u673a\u5236<\/a><\/p>\n\n\n\n

\u7684\u786e\u6709\u8f83\u5927\u63d0\u5347\u3002<\/p>\n\n\n\n

\"\"
MapTR\u4e24\u7c7b\u5730\u56fe\u5143\u7d20<\/figcaption><\/figure>\n\n\n\n
\"\"
permutation-equivalent\u63d0\u5347\u6548\u679c<\/figcaption><\/figure>\n\n\n\n
\"\"
permutation-equivalent\u56fe\u793a<\/figcaption><\/figure>\n\n\n\n

\u7b2c\u4e8c\u662f\u8bad\u7ec3\u9636\u6bb5\u5728\u5308\u7259\u5229<\/a>\u5339\u914d\u8fc7\u7a0b\u4e2d\u4e5f\u5206\u4e3a\u4e24\u7ea7,\u5148\u8fdb\u884cinstance\u7ea7\u7684\u5339\u914d,\u5339\u914d\u6210\u529f\u7684\u518d\u8fdb\u884cpoint\u7ea7\u7684\u5339\u914d,\u4e00\u662f\u66f4\u7b26\u5408\u4e8c\u7ea7query\u7684\u8bbe\u5b9a,\u7b2c\u4e8c\u4e5f\u662f\u4e3a\u4e86\u4fdd\u8bc1\u4e00\u5b9a\u7684\u987a\u5e8f(2\u79cd\u548c2k\u79cd),\u5e76\u4e14\u505a\u5230permutation-equivalent.\u800cVectorMapNet\u7531\u4e8e\u53ea\u68c0\u6d4b\u5173\u952e\u70b9,\u4e0d\u6d89\u53ca\u987a\u5e8f,\u6240\u4ee5\u53ea\u8981\u4e00\u6b21\u5339\u914d\u5c31\u53ef\u4ee5\u3002<\/p>\n\n\n\n

\u53e6\u5916MapTR\u4f7f\u7528\u4e86GKT[7]\u4f5c\u4e3a\u591a\u89c6\u89d2\u56fe\u50cf\u5f97\u5230BEV feature\u7684\u65b9\u5f0f,\u6587\u4e2d\u8bf4\u6548\u7387\u9ad8\u4e14\u5bb9\u6613\u90e8\u7f72\u3002<\/p>\n\n\n\n

[7] Efficient and Robust 2D-to-BEV Representation Learning via Geometry-guided Kernel Transformer<\/p>\n\n\n\n

\u4e0b\u9762\u662f\u4e09\u79cd\u7ecf\u5178\u65b9\u6848\u7684\u6548\u679c\u5bf9\u6bd4:<\/p>\n\n\n\n

\"\"
\u7ed3\u679c\u5bf9\u6bd4<\/figcaption><\/figure>\n\n\n\n

\u53ef\u89c1MapTR\u65e0\u8bba\u7cbe\u5ea6\u8fd8\u662f\u6548\u7387\u90fd\u6bd4\u524d\u4e24\u79cd\u6709\u663e\u8457\u7684\u63d0\u5347,\u4e14\u6574\u4e2a\u6846\u67b6\u7b80\u6d01\u4f18\u7f8e\u7aef\u5230\u7aef,\u53ef\u4ee5\u8bf4\u662f\u76ee\u524dBEV\u5728\u7ebf\u5c40\u90e8\u5730\u56fe\u9886\u57df\u975e\u5e38\u4f18\u79c0\u7684\u65b9\u6848\u4e86.\u8fd9\u4e2a\u9886\u57df\u535a\u4e3b\u5176\u5b9e\u4e5f\u662f\u521a\u521a\u63a5\u89e6,\u6ca1\u60f3\u5230\u7ed5\u6765\u7ed5\u53bb\u53c8\u7ed5\u56deDETR\u4e86,\u6709\u4e2a\u611f\u6168\u662f\u4e00\u5207\u7686\u53efBEV,\u4e00\u5207\u7686\u53efDETR,\u8fd8\u6709\u4ec0\u4e48\u662f\u4e00\u4e2a\u4f18\u79c0\u7684query\u5f97\u4e0d\u5230\u7684\u5462?\u5927\u5bb6\u6709\u5176\u4ed6\u4f18\u79c0\u89c1\u89e3\u6b22\u8fce\u7559\u8a00\u8ba8\u8bba. \u6700\u540e\u4e0a\u4e00\u5f20MapTR\u7684\u6548\u679c\u56fe.<\/p>\n\n\n\n

\"\"
mapTR\u6548\u679c<\/figcaption><\/figure>\n","protected":false},"excerpt":{"rendered":"

\u4f17\u6240\u5468\u77e5,\u7531\u4e8e\u79bb\u7ebf\u9ad8\u7cbe\u5730\u56feHigh-Defin […]<\/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-7280","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\/7280"}],"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=7280"}],"version-history":[{"count":2,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/7280\/revisions"}],"predecessor-version":[{"id":7293,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/7280\/revisions\/7293"}],"wp:attachment":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/media?parent=7280"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/categories?post=7280"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/tags?post=7280"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}