{"id":30436,"date":"2025-06-30T17:30:38","date_gmt":"2025-06-30T09:30:38","guid":{"rendered":"http:\/\/192.168.10.115\/?p=30436"},"modified":"2025-06-30T17:30:38","modified_gmt":"2025-06-30T09:30:38","slug":"2025-06-30-%e5%9c%b0%e5%b9%b3%e7%ba%bf%e9%9d%99%e6%80%81%e7%9b%ae%e6%a0%87%e6%a3%80%e6%b5%8b-maptr-%e5%8f%82%e8%80%83%e7%ae%97%e6%b3%95-v2-0","status":"publish","type":"post","link":"http:\/\/222.128.65.89:18086\/index.php\/2025\/06\/30\/30436\/","title":{"rendered":"2025-06-30 \u5730\u5e73\u7ebf\u9759\u6001\u76ee\u6807\u68c0\u6d4b MapTR \u53c2\u8003\u7b97\u6cd5 &#8211; V2.0"},"content":{"rendered":"\n<p>\u539f\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/zhuanlan.zhihu.com\/p\/1922336098576761075\">https:\/\/zhuanlan.zhihu.com\/p\/1922336098576761075<\/a><\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>\u8be5\u793a\u4f8b\u4e3a\u53c2\u8003\u7b97\u6cd5\uff0c\u4ec5\u4f5c\u4e3a\u5728\u5f81\u7a0b 6 \u4e0a\u6a21\u578b\u90e8\u7f72\u7684\u8bbe\u8ba1\u53c2\u8003\uff0c\u975e\u91cf\u4ea7\u7b97\u6cd5<\/p>\n<\/blockquote>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u7b80\u4ecb<\/h2>\n\n\n\n<p>\u9ad8\u6e05\u5730\u56fe\u662f\u81ea\u52a8\u9a7e\u9a76\u7cfb\u7edf\u7684\u91cd\u8981\u7ec4\u4ef6\uff0c\u63d0\u4f9b\u7cbe\u786e\u7684\u9a7e\u9a76\u73af\u5883\u4fe1\u606f\u548c\u9053\u8def\u8bed\u4e49\u4fe1\u606f\u3002\u4f20\u7edf\u79bb\u7ebf\u5730\u56fe\u6784\u5efa\u65b9\u6cd5\u6210\u672c\u9ad8\uff0c\u7ef4\u62a4\u590d\u6742\uff0c\u4f7f\u5f97\u4f9d\u8d56\u8f66\u8f7d\u4f20\u611f\u5668\u7684\u5b9e\u65f6\u611f\u77e5\u5efa\u56fe\u6210\u4e3a\u65b0\u8d8b\u52bf\u3002\u65e9\u671f\u5b9e\u65f6\u5efa\u56fe\u65b9\u6cd5\u5b58\u5728\u5c40\u9650\u6027\uff0c\u5982\u5904\u7406\u590d\u6742\u5730\u56fe\u5143\u7d20\u7684\u80fd\u529b\u4e0d\u8db3\u3001\u7f3a\u4e4f\u5b9e\u4f8b\u7ea7\u4fe1\u606f\u7b49\uff0c\u5728\u5b9e\u65f6\u6027\u548c\u540e\u5904\u7406\u590d\u6742\u5ea6\u4e0a\u5b58\u5728\u6311\u6218\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0c\u57fa\u4e8e <a href=\"https:\/\/zhida.zhihu.com\/search?content_id=259669073&amp;content_type=Article&amp;match_order=1&amp;q=Transformer&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">Transformer<\/a><\/p>\n\n\n\n<p>\u7684 MapTR \u6a21\u578b\u88ab\u63d0\u51fa\uff0c\u5b83\u91c7\u7528\u7aef\u5230\u7aef\u7ed3\u6784\uff0c\u4ec5\u4f7f\u7528\u56fe\u50cf\u6570\u636e\u5c31\u80fd\u5b9e\u73b0\u9ad8\u7cbe\u5ea6\u5efa\u56fe\uff0c\u540c\u65f6\u4fdd\u8bc1\u5b9e\u65f6\u6027\u548c\u9c81\u68d2\u6027\u3002MapTRv2 \u5728\u6b64\u57fa\u7840\u4e0a\u589e\u52a0\u4e86\u65b0\u7279\u6027\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u4e86\u5efa\u56fe\u7cbe\u5ea6\u548c\u6027\u80fd\u3002<\/p>\n\n\n\n<p>\u5730\u5e73\u7ebf\u9762\u5411\u667a\u9a7e\u573a\u666f\u63a8\u51fa\u7684\u5f81\u7a0b 6 \u7cfb\u5217\uff08\u5f81\u7a0b 6\uff09\u82af\u7247\uff0c\u5728\u63d0\u4f9b\u5f3a\u5927\u7b97\u529b\u7684\u540c\u65f6\u5e26\u6765\u4e86\u6781\u81f4\u7684\u6027\u4ef7\u6bd4\uff0c\u5f81\u7a0b 6 \u82af\u7247\u5bf9\u4e8e Transformer \u6a21\u578b\u7684\u9ad8\u6548\u652f\u6301\u52a9\u529b\u4e86 MapTR \u7cfb\u5217\u6a21\u578b\u7684\u7aef\u4fa7\u90e8\u7f72\u3002\u672c\u6587\u5c06\u8be6\u7ec6\u4ecb\u7ecd\u5730\u5e73\u7ebf\u7b97\u6cd5\u5de5\u5177\u94fe\u5728\u5f81\u7a0b 6 \u82af\u7247\u90e8\u7f72 MapTR \u7cfb\u5217\u6a21\u578b\u6240\u505a\u7684\u4f18\u5316\u4ee5\u53ca\u6a21\u578b\u7aef\u4fa7\u7684\u8868\u73b0\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001\u6027\u80fd\u7cbe\u5ea6\u6307\u6807<\/h2>\n\n\n\n<p>\u6a21\u578b\u914d\u7f6e\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pica.zhimg.com\/v2-29b744d9f4af83ad6810e51463542df6_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u6027\u80fd\u7cbe\u5ea6\u8868\u73b0\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic3.zhimg.com\/v2-d2ff9b1190167afb5a04a05d51e3ed2c_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>\u9884\u6d4b\u7684\u5730\u56fe\u5143\u7d20\uff1a\u201cdivider\u201d\uff0c\u201cped_crossing\u201d\uff0c\u201cboundary\u201d\uff1b<br>maptrv2_resnet50_bevformer_nuscenes \u9ed8\u8ba4\u4f7f\u7528 Lidar \u5750\u6807\u7cfb\uff0c\u548c\u516c\u7248\u4fdd\u6301\u4e00\u81f4\uff0c\u540c\u65f6\u9002\u914d ego \u5750\u6807\u7cfb\uff1bmaptroe_henet_tinym_bevformer_nuscenes \u9ed8\u8ba4\u4f7f\u7528 ego \u5750\u6807\u7cfb\uff0c\u589e\u52a0\u4e86 sdmap \u7684\u8f93\u5165\u878d\u5408\uff1b<br>\u91cf\u5316\u914d\u7f6e TopK\uff1a\u524d K \u4e2a\u91cf\u5316\u654f\u611f\u7684\u7b97\u5b50\uff1b\u9ed8\u8ba4 int8 \u6a21\u677f\uff1a\u4e0e grid \u76f8\u5173\u7b97\u5b50\u8bbe\u7f6e int16\uff0c\u5176\u4ed6\u5747\u8bbe\u7f6e int8<\/p>\n<\/blockquote>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001\u516c\u7248\u6a21\u578b\u4ecb\u7ecd<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">3.1 MapTR<\/h3>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pica.zhimg.com\/v2-00d682cf13aa83786f62c1c118ed8034_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<p>MapTR \u6a21\u578b\u7684\u9ed8\u8ba4\u8f93\u5165\u662f\u8f66\u8f7d\u6444\u50cf\u5934\u91c7\u96c6\u5230\u7684 6 \u5f20\u76f8\u540c\u5206\u8fa8\u7387\u7684\u73af\u89c6\u56fe\u50cf\uff0c\u4f7f\u7528 <a href=\"https:\/\/zhida.zhihu.com\/search?content_id=259669073&amp;content_type=Article&amp;match_order=1&amp;q=nuScenes&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">nuScenes<\/a><\/p>\n\n\n\n<p>\u6570\u636e\u96c6\uff0c\u540c\u65f6\u4e5f\u652f\u6301\u62d3\u5c55\u4e3a\u591a\u6a21\u6001\u8f93\u5165\u4f8b\u5982\u96f7\u8fbe\u70b9\u4e91\u3002\u6a21\u578b\u8f93\u51fa\u662f\u77e2\u91cf\u5316\u7684\u5730\u56fe\u5143\u7d20\u4fe1\u606f\uff0c\u5176\u4e2d\u5730\u56fe\u5143\u7d20\u4e3a\u4eba\u884c\u6a2a\u9053\u3001\u8f66\u9053\u5206\u9694\u7ebf\u548c\u9053\u8def\u8fb9\u754c 3 \u79cd\u3002\u6a21\u578b\u4e3b\u4f53\u91c7\u7528 encoder-decoder \u7684\u7aef\u5230\u7aef\u7ed3\u6784\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Map Encoder \u901a\u8fc7 CNN Backbone+BEV Encoder \u8d1f\u8d23\u63d0\u53d6 2D \u56fe\u50cf\u7279\u5f81\u5e76\u8f6c\u6362\u5230\u7edf\u4e00\u7684 BEV \u89c6\u89d2\u3002MapTR-nano \u9ed8\u8ba4\u4f7f\u7528 ResNet18 \u4f5c\u4e3a Backbone\uff0cMapTR-tiny \u9ed8\u8ba4\u4f7f\u7528 ResNet50\u3002MapTR \u517c\u5bb9\u591a\u79cd BEV Encoder \u5b9e\u73b0\u65b9\u5f0f\u4f8b\u5982 GKT\u3001LSS \u548c IPM \u7b49\u5e76\u4e14\u8868\u73b0\u7a33\u5b9a\uff0c\u9274\u4e8e GKT \u7684\u90e8\u7f72\u9ad8\u6548\u6027\u4ee5\u53ca\u5728\u6d88\u878d\u5b9e\u9a8c\u4e2d\u7684\u7cbe\u5ea6\u8868\u73b0\u66f4\u597d\uff0c\u516c\u7248 MapTR \u4f7f\u7528 GKT \u4f5c\u4e3a\u9ed8\u8ba4 BEV Encoder \u5b9e\u73b0\u65b9\u5f0f\u3002<\/li>\n\n\n\n<li>Map Decoder \u91c7\u7528 Hierarchical Query Embedding Scheme\uff0c\u5373\u4ece point-level\uff08\u4f4d\u7f6e\uff09\u548c instance-level\uff08\u8f6e\u5ed3\uff09\u663e\u5f0f\u5730\u7f16\u7801\u5730\u56fe\u5143\u7d20\uff0cpoint-level queries \u88ab\u6240\u6709 instances \u5171\u4eab\u5e76\u878d\u5408\u8fdb instance-level queries \u4ece\u800c\u751f\u6210 hierarchical queries\uff0chierarchical queries \u7ecf\u8fc7\u7ea7\u8054\u7684 decoder layers\uff08\u9ed8\u8ba4\u662f 6 \u5c42\uff09\u4e0d\u65ad\u66f4\u65b0\u3002\u6bcf\u4e2a decoder layer \u9996\u5148\u4f7f\u7528\u591a\u5934\u81ea\u6ce8\u610f\u529b\uff08MHSA\uff09\u505a inter-instance \u548c intra-instance \u7684\u4fe1\u606f\u4ea4\u4e92\uff0c\u63a5\u7740\u4f1a\u4f7f\u7528 <a href=\"https:\/\/zhida.zhihu.com\/search?content_id=259669073&amp;content_type=Article&amp;match_order=1&amp;q=Deformable+Attention&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">Deformable Attention<\/a><\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6765\u4e0e Map Encoder \u8f93\u51fa\u7684 BEV \u7279\u5f81\u505a\u4fe1\u606f\u4ea4\u4e92\u3002point-level \u7684\u4fe1\u606f\u88ab\u6240\u6709 instance \u5171\u4eab\uff0c\u6240\u4ee5\u5bf9\u4e8e\u6bcf\u4e2a instance \u800c\u8a00\uff0c\u6620\u5c04\u5230 BEV \u7a7a\u95f4\u7684\u591a\u4e2a\u53c2\u8003\u70b9 reference points \u662f\u7075\u6d3b\u4e14\u52a8\u6001\u5206\u5e03\u7684\uff0c\u8fd9\u5bf9\u4e8e\u63d0\u53d6 long-range context information \u9884\u6d4b\u968f\u673a\u5f62\u72b6\u7684\u5730\u56fe\u5143\u7d20\u662f\u6709\u76ca\u7684\u3002<\/li>\n\n\n\n<li>MapTR Head \u7531\u5206\u7c7b\u5206\u652f\u548c\u56de\u5f52\u5206\u652f\u6784\u6210\u3002\u5206\u7c7b\u5206\u652f\u9884\u6d4b instances \u7684\u7c7b\u522b\uff0c\u56de\u5f52\u5206\u652f\u9884\u6d4b points \u96c6\u5408\u7684\u4f4d\u7f6e\u3002Head \u8f93\u51fa\u7684\u9884\u6d4b\u503c\u548c\u771f\u503c GT \u4e4b\u95f4\u91c7\u7528 Hierarchical Bipartite Matching \u5b9e\u73b0\u76d1\u7763\u5b66\u4e60\uff0c\u5206\u4e3a Instance-level Matching \u548c Point-level Matching\uff0c\u56e0\u6b64\u635f\u5931\u51fd\u6570\u4e3a\u4e09\u4e2a\u90e8\u5206\u7684\u52a0\u6743\u548c\uff1a\u5206\u7c7b Classification Loss\u3001\u70b9\u5bf9\u70b9\u4f4d\u7f6e Point2point Loss \u548c\u8fde\u63a5\u8fb9\u65b9\u5411 Edge Direction Loss\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">3.2 MapTRv2<\/h3>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic2.zhimg.com\/v2-1a9f929ec35d1030d1b4c725ffbb8f45_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<p>MapTRv2 \u5728 MapTR \u7684\u57fa\u7840\u4e0a\u589e\u52a0\u4e86\u65b0\u7684\u7279\u6027\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u9488\u5bf9\u5c42\u6b21\u5316 query\uff0c\u5f15\u5165\u89e3\u8026\u81ea\u6ce8\u610f\u529b\uff0c\u6781\u5927\u5730\u51cf\u5c11\u4e86\u8ba1\u7b97\u91cf\u548c\u663e\u5b58\u6d88\u8017\uff1b\u5bf9\u4e8e\u548c\u8f93\u5165\u7279\u5f81\u4ea4\u4e92\u7684 cross-attention \u90e8\u5206\uff0c\u5219\u5f15\u5165\u4e86 BEV\u3001PV \u548c BEV+PV \u4e09\u79cd\u53d8\u4f53\uff1b<\/li>\n\n\n\n<li>\u5f15\u5165\u8f85\u52a9 one-to-many \u96c6\u5408\u9884\u6d4b\u5206\u652f\uff0c\u589e\u52a0\u4e86\u6b63\u6837\u672c\u6570\uff0c\u52a0\u901f\u4e86\u8bad\u7ec3\u6536\u655b\uff1b<\/li>\n\n\n\n<li>\u5f15\u5165\u8f85\u52a9 dense supervision\uff0c\u5f15\u5165\u6df1\u5ea6\u4f30\u8ba1\u9884\u6d4b\u5934\u3001PV \u548c BEV \u89c6\u89d2\u4e0b\u7684\u5206\u5272\u5934\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u6a21\u578b\u7cbe\u5ea6\u3002\u7531\u4e8e\u5f15\u5165\u6df1\u5ea6\u4fe1\u606f\u505a\u76d1\u7763\u5b66\u4e60\uff0c\u4e3a\u4e86\u663e\u5f0f\u5730\u63d0\u53d6\u6df1\u5ea6\u4fe1\u606f\uff0c\u516c\u7248 MapTRv2 \u9009\u62e9\u57fa\u4e8e LSS \u7684 <a href=\"https:\/\/zhida.zhihu.com\/search?content_id=259669073&amp;content_type=Article&amp;match_order=1&amp;q=BEVPoolv2&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">BEVPoolv2<\/a><\/li>\n<\/ol>\n\n\n\n<p>\u6765\u4f5c\u4e3a BEV \u89c6\u89d2\u8f6c\u6362\u65b9\u5f0f\uff1b\u5f15\u5165\u65b0\u7684\u5730\u56fe\u5143\u7d20\u8f66\u9053\u4e2d\u5fc3\u7ebf\uff08centerline\uff09\uff1b\u589e\u52a0 3D \u5730\u56fe\u5143\u7d20\u9884\u6d4b\u80fd\u529b\uff0c\u5e76\u63d0\u4f9b <a href=\"https:\/\/zhida.zhihu.com\/search?content_id=259669073&amp;content_type=Article&amp;match_order=1&amp;q=Argoverse2&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">Argoverse2<\/a><\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u6570\u636e\u96c6\u4e0a\u7684\u6307\u6807\u3002<\/li>\n<\/ol>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001\u5730\u5e73\u7ebf\u90e8\u7f72\u8bf4\u660e<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>\u5730\u5e73\u7ebf\u53c2\u8003\u7b97\u6cd5\u4f7f\u7528\u6d41\u7a0b\u8bf7\u53c2\u8003<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/developer.horizon.auto\/bloggerdetail%3Fbid%3D654209252113248256\" target=\"_blank\" rel=\"noreferrer noopener\">\u5f81\u7a0b 6 \u53c2\u8003\u7b97\u6cd5\u4f7f\u7528\u6307\u5357<\/a>\uff1b\u5bf9\u5e94\u9ad8\u6548\u6a21\u578b\u8bbe\u8ba1\u5efa\u8bae\u8bf7\u53c2\u8003\u300a\u5f81\u7a0b 6 \u5e73\u53f0\u7b97\u6cd5\u8bbe\u8ba1\u5efa\u8bae\u300b<\/p>\n<\/blockquote>\n\n\n\n<p>MapTROE \u6a21\u578b\u5f15\u5165\u4e86 SD map \u7684\u524d\u878d\u5408\u7ed3\u6784\uff0c\u4e0e\u56fe\u50cf\u89c6\u89d2\u8f6c\u6362\u540e\u7684 bev feature \u8fdb\u884c\u878d\u5408\uff0c\u518d\u901a\u8fc7\u4f18\u5316\u540e\u7684 MapTR head \u751f\u6210\u77e2\u91cf\u5316\u7684\u5730\u56fe\u5143\u7d20\u3002\u6574\u4f53\u7ed3\u6784\u5982\u4e0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic4.zhimg.com\/v2-1ae3ed2724e1fb014703b78b48c78441_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u56e0\u6b64 maptroe_henet_tinym_bevformer_nuscenes \u6a21\u578b\u76f8\u6bd4\u4e4b\u524d\u7248\u672c\u65b0\u589e\u4e86\u5982\u4e0b\u4f18\u5316\u70b9\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5f15\u5165 SD map \u524d\u878d\u5408\uff0c\u63d0\u5347\u6a21\u578b\u6574\u4f53\u7cbe\u5ea6\u8868\u73b0\uff08\u7cbe\u5ea6\u6307\u6807\u63d0\u5347 7+\uff09\uff1b<\/li>\n\n\n\n<li>MapTR Head \u90e8\u5206\u4f18\u5316\u4e3a Instance Head\uff0c\u5728\u7cbe\u5ea6\u76f8\u5f53\u7684\u60c5\u51b5\u4e0b\u6027\u80fd\u63d0\u5347 35%\uff1b<\/li>\n\n\n\n<li>\u91c7\u7528\u5f81\u7a0b 6 \u82af\u7247\u9ad8\u6548 backbone HENet_tinym\uff0c\u5728\u7cbe\u5ea6\u8f7b\u5fae\u63d0\u5347\u7684\u60c5\u51b5\u4e0b\u6781\u5927\u63d0\u9ad8\u4e86\u6027\u80fd\uff1b<\/li>\n\n\n\n<li>View Transformer \u91c7\u7528\u4f18\u5316\u7248 bevformer\uff0c\u5728\u7cbe\u5ea6\u76f8\u5f53\u7684\u60c5\u51b5\u4e0b\u63d0\u9ad8\u4e86\u6027\u80fd\u3002<\/li>\n<\/ul>\n\n\n\n<p>maptroe_henet_tinym_bevformer_nuscenes \u6a21\u578b\u5bf9\u5e94\u7684\u4ee3\u7801\u8def\u5f84\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/picx.zhimg.com\/v2-18bb1a0f4088265120263ef7703a6121_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 \u6027\u80fd\u4f18\u5316<\/h3>\n\n\n\n<h3 class=\"wp-block-heading\">4.1.1 Backbone<\/h3>\n\n\n\n<p>MapTROE \u91c7\u7528\u57fa\u4e8e\u5f81\u7a0b 6 \u82af\u7247\u7684\u9ad8\u6548\u8f7b\u91cf\u5316 Backbone <a href=\"https:\/\/zhida.zhihu.com\/search?content_id=259669073&amp;content_type=Article&amp;match_order=1&amp;q=HENet_TinyM&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">HENet_TinyM<\/a><\/p>\n\n\n\n<p>\uff08Hybrid Efficient Network\uff0c Tiny for J6M\uff09\uff0cHENet \u80fd\u66f4\u597d\u5730\u5229\u7528\u5f81\u7a0b 6 \u7cfb\u5217\u82af\u7247\u7684\u7b97\u529b\uff0c\u5728\u6a21\u578b\u7cbe\u5ea6\u548c\u6027\u80fd\u4e0a\u66f4\u5177\u4f18\u52bf\u3002HENet_TinyM \u91c7\u7528\u4e86\u7eaf CNN \u67b6\u6784\uff0c\u603b\u4f53\u5206\u4e3a\u56db\u4e2a stage\uff0c\u6bcf\u4e2a stage \u4f1a\u8fdb\u884c\u4e00\u6b21 2 \u500d\u4e0b\u91c7\u6837\uff0c\u5177\u4f53\u7ed3\u6784\u914d\u7f6e\u5982\u4e0b\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># henet-tinym\ndepth = &#91;4, 3, 8, 6]\nblock_cls = &#91;\"GroupDWCB\", \"GroupDWCB\", \"AltDWCB\", \"DWCB\"]\nwidth = &#91;64, 128, 192, 384]\nattention_block_num = &#91;0, 0, 0, 0]\nmlp_ratios, mlp_ratio_attn = &#91;2, 2, 2, 3], 2\nact_layer = &#91;\"nn.GELU\", \"nn.GELU\", \"nn.GELU\", \"nn.GELU\"]\nuse_layer_scale = &#91;True, True, True, True]\nextra_act = &#91;False, False, False, False]\nfinal_expand_channel, feature_mix_channel = 0, 1024\ndown_cls = &#91;\"S2DDown\", \"S2DDown\", \"S2DDown\", \"None\"]\npatch_embed = \"origin\"<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">4.1.2 Neck<\/h3>\n\n\n\n<p>Neck \u90e8\u5206\u91c7\u7528\u4e86\u5730\u5e73\u7ebf\u5185\u90e8\u5b9e\u73b0\u7684 FPN\uff0c\u76f8\u6bd4\u516c\u7248 FPN \u5b9e\u73b0\uff0c\u5728\u5f81\u7a0b 6 \u5e73\u53f0\u4e0a\u6027\u80fd\u66f4\u52a0\u53cb\u597d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.1.3 View Transformer<\/h3>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>\u5730\u5e73\u7ebf\u53c2\u8003\u7b97\u6cd5\u7248\u672c\u5c06\u57fa\u4e8e LSS \u7684\u89c6\u89d2\u8f6c\u6362\u65b9\u5f0f\u66ff\u6362\u4e3a\u6df1\u5ea6\u4f18\u5316\u540e Bevformer \u7684 View Transformer \u90e8\u5206\u3002<\/p>\n<\/blockquote>\n\n\n\n<ol class=\"wp-block-list\">\n<li>BEV Grid \u5c3a\u5bf8\uff1a\u5bf9\u4e8e Dense BEV \u800c\u8a00\uff0cBEV Grid \u7684\u5c3a\u5bf8\u5927\u5c0f\u5b9e\u9645\u5730\u5f71\u54cd\u6a21\u578b\u6027\u80fd\u3002\u5f81\u7a0b 6 \u5e73\u53f0\u589e\u5f3a\u4e86\u5e26\u5bbd\u80fd\u529b\uff0c\u4f46\u4ecd\u9700\u6ce8\u610f BEV \u7f51\u683c\u8fc7\u5927\u5bfc\u81f4\u8bbf\u5b58\u538b\u529b\u8fc7\u5927\u800c\u5bf9\u6027\u80fd\u5e26\u6765\u8d1f\u9762\u5f71\u54cd\uff0c\u5efa\u8bae\u8003\u8651\u5b9e\u9645\u90e8\u7f72\u60c5\u51b5\u9009\u62e9\u5408\u9002\u7684 BEV \u7f51\u683c\u5927\u5c0f\u6765\u8bbe\u8ba1\u6a21\u578b\u3002\u76f8\u6bd4\u516c\u7248 MapTRv2 \u6a21\u578b\u4f7f\u7528 200&#215;100 \u7684\u7f51\u683c\uff0c\u5730\u5e73\u7ebf\u90e8\u7f72\u6a21\u578b\u4f7f\u7528 100&#215;50 \u7684\u7f51\u683c\u6765\u5b9e\u73b0\u6027\u80fd\u548c\u7cbe\u5ea6\u7684\u5e73\u8861\u3002<\/li>\n\n\n\n<li>BEV \u7279\u5f81\u7f16\u7801\uff1a a\u3002 \u9ed8\u8ba4 prev_bev \u7531 cur_bev \u6539\u4e3a\u5168 0\uff1b b. \u53d6\u6d88 can_bus \u4fe1\u606f\u7684\u4f7f\u7528\uff0c\u524d\u4e00\u5e27 bev \u7279\u5f81 prev_bev \u548c\u5f53\u524d\u5e27 cur_bev \u7684\u5bf9\u9f50\u65b9\u5f0f\u7531\u4f7f\u7528 can_bus \u4fe1\u606f\u6b63\u5411\u6821\u51c6\u6539\u4e3a\u4f7f\u7528 GridSample \u7b97\u5b50\u53cd\u5411\u91c7\u6837\u6821\u51c6\uff1b c. \u53d6\u6d88\u4e86 bev_query \u521d\u59cb\u5316\u90e8\u5206\u548c can_bus \u7684\u878d\u5408\uff1b<\/li>\n<\/ol>\n\n\n\n<pre class=\"wp-block-code\"><code># \u516c\u7248\u6a21\u578b\nclass MapTRPerceptionTransformer(BaseModule):\n    ...\n    def attn_bev_encode(...):\n        ...\n        if prev_bev is not None:\n            if prev_bev.shape&#91;1] == bev_h * bev_w:\n                prev_bev = prev_bev.permute(1, 0, 2)\n            if self.rotate_prev_bev:\n                for i in range(bs):\n                    # num_prev_bev = prev_bev.size(1)\n                    rotation_angle = kwargs&#91;'img_metas']&#91;i]&#91;'can_bus']&#91;-1]\n                    tmp_prev_bev = prev_bev&#91;:, i].reshape(\n                        bev_h, bev_w, -1).permute(2, 0, 1)\n                    tmp_prev_bev = rotate(tmp_prev_bev, rotation_angle,\n                                          center=self.rotate_center)\n                    tmp_prev_bev = tmp_prev_bev.permute(1, 2, 0).reshape(\n                        bev_h * bev_w, 1, -1)\n                    prev_bev&#91;:, i] = tmp_prev_bev&#91;:, 0]\n\n        # add can bus signals\n        can_bus = bev_queries.new_tensor(\n            &#91;each&#91;'can_bus'] for each in kwargs&#91;'img_metas']])  # &#91;:, :]\n        can_bus = self.can_bus_mlp(can_bus&#91;:, :self.len_can_bus])&#91;None, :, :]\n        bev_queries = bev_queries + can_bus * self.use_can_bus\n        ...\n\n# \u5730\u5e73\u7ebf\u53c2\u8003\u7b97\u6cd5\nclass BevFormerViewTransformer(nn.Module):\n    ...\n    def __init__(...):\n        ...\n        self.prev_frame_info = {\n            \"prev_bev\": None,\n            \"scene_token\": None,\n            \"ego2global\": None,\n        }\n        ...\n    def get_prev_bev(...):\n        if idx == self.queue_length - 1 and self.queue_length != 1:\n            prev_bev = torch.zeros(\n                (bs, self.bev_h * self.bev_w, self.embed_dims),\n                dtype=torch.float32,\n                device=device,\n            )\n            ...\n        else:\n            prev_bev = self.prev_frame_info&#91;\"prev_bev\"]\n            if prev_bev is None:\n                prev_bev = torch.zeros(\n                    (bs, self.bev_h * self.bev_w, self.embed_dims),\n                    dtype=torch.float32,\n                    device=device,\n                ) # \u5bf9\u5e94\u6539\u52a82.a\n                ...\n    def bev_encoder(...):\n        ...\n        tmp_prev_bev = prev_bev.reshape(\n            bs, self.bev_h, self.bev_w, self.embed_dims\n        ).permute(0, 3, 1, 2)\n        prev_bev = F.grid_sample(\n            tmp_prev_bev, norm_coords, \"bilinear\", \"zeros\", True\n        ) # \u5bf9\u5e94\u6539\u52a82.b\n        ...\nclass SingleBevFormerViewTransformer(BevFormerViewTransformer):\n    ...\n    def get_bev_embed(...):\n        ...\n        bev_query = self.bev_embedding.weight\n        bev_query = bev_query.unsqueeze(1).repeat(1, bs, 1) # \u5bf9\u5e94\u6539\u52a82.c\n        ...<\/code><\/pre>\n\n\n\n<p>d. \u53d6\u6d88\u4e86\u516c\u7248\u7684 TemporalSelfAttention\uff0c\u6539\u4e3a HorizonMSDeformableAttention\uff0c\u4fdd\u6301\u7cbe\u5ea6\u7684\u540c\u65f6\u63d0\u5347\u901f\u5ea6\uff1b<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># \u516c\u7248\u6a21\u578bConfig\nmodel = dict(\n    ...\n    pts_bbox_head=dict(\n        type='MapTRHead',\n        ...\n        transformer=dict(\n            type='MapTRPerceptionTransformer',\n            ...\n            encoder=dict(\n                type='BEVFormerEncoder',\n                ...\n                transformerlayers=dict(\n                    type='BEVFormerLayer',\n                    attn_cfgs=&#91;\n                        dict(\n                            type='TemporalSelfAttention',\n                            embed_dims=_dim_,\n                            num_levels=1),\n                            ...\n                    ]\n                )\n            )\n        )\n    )\n)\n\n# \u5730\u5e73\u7ebf\u53c2\u8003\u7b97\u6cd5Config\nmodel = dict(\n    ...\n    view_transformer=dict(\n        type=\"SingleBevFormerViewTransformer\",\n        ...\n        encoder=dict(\n            type=\"SingleBEVFormerEncoder\",\n            ...\n            encoder_layer=dict(\n                type=\"SingleBEVFormerEncoderLayer\",\n                ...\n                selfattention=dict(\n                    type=\"HorizonMSDeformableAttention\", # \u5bf9\u5e94\u6539\u52a82.d\n                    ...\n                ),\n            )\n        )\n    )\n)<\/code><\/pre>\n\n\n\n<p>e. \u652f\u6301\u516c\u7248 Bevformer \u4e2d\u7684 bev_mask\uff0c\u5e76\u5c06\u6d89\u53ca\u5230\u7684 gather\/scatter \u64cd\u4f5c\uff0c\u7528 gridsample \u7b49\u4ef7\u66ff\u6362\uff0c\u63d0\u9ad8\u6a21\u578b\u901f\u5ea6\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># \u5730\u5e73\u7ebf\u53c2\u8003\u7b97\u6cd5Config\nview_transformer=dict(\n    type=\"SingleBevFormerViewTransformer\",\n    ...\n    max_camoverlap_num=2, # \u5bf9\u5e94\u6839\u636ebev_mask\u8fdb\u884c\u7a00\u758f\u6620\u5c04\uff0c\u63d0\u9ad8\u8fd0\u884c\u6548\u7387\uff0c\u5bf9\u5e94\u6539\u52a82.e\n    virtual_bev_h=int(0.4 * bev_h_),\n    virtual_bev_w=bev_w_,\n    ...\n)<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">4.1.4 Head<\/h3>\n\n\n\n<p>\u516c\u7248 MapTR \u4f7f\u7528\u5206\u5c42 query \u673a\u5236\uff0c\u5b9a\u4e49\u4e00\u7ec4 instance queries \u548c\u7531\u6240\u6709 instance \u5171\u4eab\u7684 point queries\uff0c\u6bcf\u4e2a\u5730\u56fe\u5143\u7d20\u5bf9\u5e94\u4e00\u7ec4\u5206\u5c42 query\uff08\u4e00\u4e2a instance query \u548c\u5171\u4eab\u7684 point queries \u5e7f\u64ad\u76f8\u52a0\u5f97\u5230\uff09\uff0c\u5728 decoder layer \u4e2d\u5206\u522b\u4f7f\u7528 self-attention \u548c cross-attention \u6765\u66f4\u65b0\u5206\u5c42 query\u3002<\/p>\n\n\n\n<p>MapTROE \u7684\u6539\u8fdb\u5219\u662f\u4e3a\u6bcf\u4e2a\u5730\u56fe\u5143\u7d20\u5206\u914d\u4e00\u4e2a instance query\uff08\u65e0\u76f4\u63a5 point query\uff09\uff0c\u6bcf\u4e2a query \u7528\u4e8e\u7f16\u7801\u8bed\u4e49\u4fe1\u606f\u548c\u5730\u7406\u4f4d\u7f6e\u4fe1\u606f\uff0cdecoder \u9636\u6bb5\u548c\u516c\u7248 MapTR \u4e00\u6837\uff0c\u5206\u522b\u8fdb\u884c multi-head self-attention \u548c deformable cross-attention\uff0c\u6700\u540e\u6bcf\u4e2a instance query \u901a\u8fc7 MLP \u7f51\u7edc\u751f\u6210\u7c7b\u522b\u4fe1\u606f\u548c\u5143\u7d20\u5185\u7684\u70b9\u96c6\u5750\u6807\uff0c\u76f8\u6bd4\u516c\u7248\u9884\u6d4b\u5206\u5c42 query\uff0c\u6539\u8fdb\u540e\u76f4\u63a5\u9884\u6d4b instance query \u5e26\u6765\u7684\u8ba1\u7b97\u91cf\u66f4\u5c11\uff0c\u6781\u5927\u5730\u63d0\u9ad8\u4e86\u6a21\u578b\u5728\u7aef\u4fa7\u7684\u8fd0\u884c\u6027\u80fd\u3002\u540c\u65f6\u501f\u9274 StreamMapNet\uff0c\u4f7f\u7528\u591a\u70b9\u6ce8\u610f\u529b\u65b9\u6cd5\u6765\u9002\u5e94\u9ad8\u5ea6\u4e0d\u89c4\u5219\u7684\u5730\u56fe\u5143\u7d20\uff0c\u6269\u5927\u611f\u77e5\u8303\u56f4\u3002\u4ee3\u7801\u89c1<code>\/usr\/local\/lib\/python3.10\/dist-packages\/hat\/models\/task_modules\/maptr\/instance_decoder.py: class MapInstanceDetectorHead(nn.Module)<\/code><\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.1.5 \u591a\u70b9\u6ce8\u610f\u529b<\/h3>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic4.zhimg.com\/v2-05e627851a77e9817cf8a22db815fa1b_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u4f20\u7edf\u7684\u53ef\u53d8\u5f62\u6ce8\u610f\u529b\u4e3a\u6bcf\u4e2a query \u5206\u914d\u4e00\u4e2a\u53c2\u8003\u70b9\uff0c\u591a\u70b9\u6ce8\u610f\u529b\u5219\u4f7f\u7528\u524d\u4e00\u5c42\u9884\u6d4b\u7684\u5730\u56fe\u5143\u7d20\u7684\u591a\u4e2a\u70b9\u4f5c\u4e3a\u5f53\u524d\u5c42 query \u7684\u53c2\u8003\u70b9\uff0c\u5177\u4f53\u8ba1\u7b97\u65b9\u5f0f\u662f\u5728\u70b9\u7ef4\u5ea6\u4e0a\u6269\u5c55\u4e86\u4e00\u5c42\u6c42\u548c\uff0c\u5c06\u4e00\u4e2a\u70b9\u53d8\u6210\u591a\u4e2a\u70b9\uff0c\u5206\u522b\u8ba1\u7b97 deformable attention\u3002\u56de\u5f52\u7684\u65f6\u5019\u5e76\u975e\u9884\u6d4b offsets\uff0c\u800c\u662f\u76f4\u63a5\u9884\u6d4b\u5730\u56fe\u5143\u7d20\u70b9\u7684\u5750\u6807\u4f4d\u7f6e\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.1.6Attention<\/h3>\n\n\n\n<p>\u6a21\u578b\u4e2d\u7528\u5230\u7684 attention \u64cd\u4f5c\u5747\u4f7f\u7528\u5730\u5e73\u7ebf\u63d0\u4f9b\u7684\u7b97\u5b50\uff0c\u76f8\u6bd4 PyTorch \u63d0\u4f9b\u7684\u516c\u7248\u7b97\u5b50\uff0c\u5730\u5e73\u7ebf attention \u7b97\u5b50\u5728\u4fdd\u6301\u7b97\u5b50\u903b\u8f91\u7b49\u4ef7\u7684\u540c\u65f6\u5728\u6548\u7387\u4e0a\u8fdb\u884c\u4e86\u4f18\u5316<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>from hat.models.task_modules.bevformer.attention import (\n    HorizonMSDeformableAttention,\n    HorizonMSDeformableAttention3D,\n    HorizonSpatialCrossAttention,\n    ...\n)<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 \u7cbe\u5ea6\u4f18\u5316<\/h3>\n\n\n\n<h3 class=\"wp-block-heading\">4.2.1 \u6d6e\u70b9\u7cbe\u5ea6<\/h3>\n\n\n\n<p>MapTROE \u6a21\u578b\u5f15\u5165 SD Map \u524d\u878d\u5408\uff0c\u4e0e\u56fe\u50cf\u8f6c\u6362\u540e\u7684 bev feature \u8fdb\u884c\u878d\u5408\uff0c\u4ee5\u63d0\u9ad8\u5728\u7ebf\u5730\u56fe\u7684\u751f\u6210\u8d28\u91cf\u3002\u6a21\u5757\u7ed3\u6784\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic1.zhimg.com\/v2-783a7ef0da41381782751c37ff142456_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">4.2.1.1 SD Map \u7279\u5f81\u63d0\u53d6<\/h3>\n\n\n\n<p>SD Map \u4ece OpenStreetMap\uff08OSM\uff09\u4e2d\u83b7\u53d6\uff0c\u901a\u8fc7\u7531 GPS \u63d0\u4f9b\u7684\u8f66\u8f86\u4f4d\u59ff\uff0c\u67e5\u8be2\u8f66\u8f86\u5f53\u524d\u4f4d\u59ff\u9644\u8fd1\u7684 SD Map\uff0c\u7136\u540e\u5c06 SD Map \u8f6c\u6362\u5230\u81ea\u8f66\u5750\u6807\u7cfb\u4e0b\uff0c\u4e0e NuScenes \u4e2d\u7684\u6570\u636e\u6807\u6ce8\u5750\u6807\u7cfb\u4fdd\u6301\u4e00\u81f4\u3002SD Map \u4f1a\u4ece\u8f66\u9053\u4e2d\u5fc3\u9aa8\u67b6\u7ebf Polyline \u7684\u5f62\u5f0f\u8f6c\u5316\u4e3a\u6805\u683c\u7ed3\u6784\uff0c\u5927\u5c0f\u548c BEV \u7279\u5f81\u76f8\u540c\uff0c\u7ecf\u8fc7 CNN \u53d8\u6210\u7279\u5f81\u56fe\uff0c\u5bf9\u5e94 SD Map \u7684\u5148\u9a8c\u4fe1\u606f\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.2.1.2 SD Map \u7279\u5f81\u878d\u5408<\/h3>\n\n\n\n<p>\u6805\u683c\u5316\u540e\u7684 SD Map \u548c\u5b9e\u9645\u573a\u666f\u53ef\u80fd\u4f1a\u51fa\u73b0\u9519\u4f4d\u3001\u4e0d\u5bf9\u9f50\u7684\u60c5\u51b5\uff0c\u8fd9\u79cd\u9519\u4f4d\u5bfc\u81f4\u76f4\u63a5 Concatenate BEV \u7279\u5f81\u548c SD Map \u7279\u5f81\u7684\u6548\u679c\u5e76\u4e0d\u597d\uff0c\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u5f15\u5165\u4e86\u7279\u5f81\u878d\u5408\u6a21\u5757\uff0c\u901a\u8fc7\u7f51\u7edc\u5b66\u4e60\u6765\u51b3\u5b9a\u6700\u9002\u5408\u7684\u5bf9\u9f50\u65b9\u5f0f\uff0c\u53ef\u4ee5\u6709\u6548\u5730\u5229\u7528 SD Map \u5148\u9a8c\u63d0\u5347 BEV \u7279\u5f81\u7684\u6548\u679c\u3002\u5173\u4e8e\u7279\u5f81\u878d\u5408\u6a21\u5757\uff0c\u5206\u522b\u5b9e\u9a8c\u4e86\u4ea4\u53c9\u6ce8\u610f\u529b\u4e0e CNN \u7f51\u7edc\uff0c\u901a\u8fc7\u7cbe\u5ea6\u4e0e\u6027\u80fd\u7684\u5e73\u8861\uff0c\u6700\u540e\u9009\u62e9\u4e86 CNN \u7f51\u7edc\u6a21\u5757\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.3 \u91cf\u5316\u7cbe\u5ea6<\/h3>\n\n\n\n<ol class=\"wp-block-list\">\n<li>maptroe_henet_tinym_bevformer_nuscenes \u6a21\u578b\u9ed8\u8ba4\u4f7f\u7528 int8 \u6a21\u677f\uff0c\u4e5f\u5373\u4e0e grid \u76f8\u5173\u7b97\u5b50\u8bbe\u7f6e int16\uff0c\u5176\u4ed6\u5747\u8bbe\u7f6e int8\uff0c\u76f8\u6bd4 maptrv2_resnet50_bevformer_nuscenes \u7684\u91cf\u5316\u914d\u7f6e\uff0cint16 \u91cf\u5316\u7684\u7b97\u5b50\u66f4\u5c11\uff0c\u6a21\u578b\u7aef\u4fa7\u6027\u80fd\u66f4\u9ad8<\/li>\n<\/ol>\n\n\n\n<pre class=\"wp-block-code\"><code># Config\u6587\u4ef6\ncali_qconfig_setter = (default_calibration_qconfig_setter,)\nqat_qconfig_setter = (default_qat_fixed_act_qconfig_setter,)<\/code><\/pre>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u6d6e\u70b9\u9636\u6bb5\u91c7\u7528\u66f4\u5927\u7684 weight decay \u8bad\u7ec3\uff0c\u4f7f\u6d6e\u70b9\u6570\u636e\u5206\u5e03\u8303\u56f4\u66f4\u5c0f\uff0c\u6d6e\u70b9\u6a21\u578b\u53c2\u6570\u66f4\u6709\u5229\u4e8e\u91cf\u5316<\/li>\n<\/ol>\n\n\n\n<pre class=\"wp-block-code\"><code># Config\u6587\u4ef6\nfloat_trainer = dict(\n    ...\n    optimizer=dict(\n        ...\n        weight_decay=0.1, # \u76f8\u6bd4maptrv2_resnet50_bevformer_nuscenes\u589e\u5927\u4e8610\u500d\n    ),\n    ...\n)<\/code><\/pre>\n\n\n\n<ol class=\"wp-block-list\">\n<li>QAT \u8bad\u7ec3\u91c7\u7528\u56fa\u5b9a\u8f83\u5c0f\u7684 learning rate \u6765 fine-tune\uff0c\u8fd9\u91cc\u56fa\u5b9a\u4e5f\u5373\u53d6\u6d88 LrUpdater Callback \u7684\u4f7f\u7528\uff0c\u914d\u7f6e\u5982\u4e0b\uff1a<\/li>\n<\/ol>\n\n\n\n<pre class=\"wp-block-code\"><code># Config\u6587\u4ef6\nqat_lr = 1e-9<\/code><\/pre>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u53d6\u6d88\u4e86\u516c\u7248\u6a21\u578b MapTRHead \u4e2d\u5bf9\u4e8e\u91cf\u5316\u4e0d\u53cb\u597d\u7684 inverse_sigmoid \u64cd\u4f5c\uff1b\u6b64\u5916 MapTROE \u5bf9 Head \u7684\u4f18\u5316\u65e0\u9700\u518d\u5f15\u5165 reg_branches \u8f93\u51fa\u548c reference \u76f8\u52a0\u540e\u518d sigmoid \u7684\u64cd\u4f5c\uff1a<\/li>\n<\/ol>\n\n\n\n<pre class=\"wp-block-code\"><code># \u516c\u7248\u6a21\u578b\nclass MapTRHead(DETRHead):\n    ...\n    def forward(...):\n        ...\n        for lvl in range(hs.shape&#91;0]):\n            if lvl == 0:\n                # import pdb;pdb.set_trace()\n                reference = init_reference\n            else:\n                reference = inter_references&#91;lvl - 1]\n            reference = inverse_sigmoid(reference)\n            ...\n            tmp = self.reg_branches&#91;lvl](...)\n            tmp&#91;..., 0:2] += reference&#91;..., 0:2]\n            tmp = tmp.sigmoid() # cx,cy,w,h\n\n# \u5730\u5e73\u7ebf\u53c2\u8003\u7b97\u6cd5\nclass MapInstanceDetectorHead(nn.Module):\n    ...\n    def get_outputs(...):\n        ...\n        for lvl in range(len(outputs_classes)):\n            tmp = reference_out&#91;lvl].float()\n\n            outputs_coord, outputs_pts_coord = self.transform_box(tmp)\n            outputs_class = outputs_classes&#91;lvl].float()\n\n            outputs_classes_one2one.append(\n                outputs_class&#91;:, 0 : self.num_vec_one2one]\n            )\n            outputs_coords_one2one.append(\n                outputs_coord&#91;:, 0 : self.num_vec_one2one]\n            )\n            outputs_pts_coords_one2one.append(\n                outputs_pts_coord&#91;:, 0 : self.num_vec_one2one]\n            )\n\n            outputs_classes_one2many.append(\n                outputs_class&#91;:, self.num_vec_one2one :]\n            )\n            outputs_coords_one2many.append(\n                outputs_coord&#91;:, self.num_vec_one2one :]\n            )\n            outputs_pts_coords_one2many.append(\n                outputs_pts_coord&#91;:, self.num_vec_one2one :]\n            )\n    ...        \n    def forward(...):\n        outputs = self.bev_decoder(...)\n        if self.is_deploy:\n            return outputs\n        ...\n        outputs = self.get_outputs(...)\n        ...\n        return self._post_process(data, outputs)<\/code><\/pre>\n\n\n\n<ol class=\"wp-block-list\">\n<li>Attention \u7ed3\u6784\u4f18\u5316\uff0c\u901a\u8fc7\u6570\u503c\u878d\u5408\u65b9\u6cd5\uff0c\u5c06\u90e8\u5206\u6570\u503c\u8fd0\u7b97\u63d0\u524d\u8fdb\u884c\u878d\u5408\uff0c\u51cf\u5c11\u6574\u4f53\u7684\u91cf\u5316\u64cd\u4f5c\uff0c\u63d0\u9ad8\u6a21\u578b\u7684\u91cf\u5316\u53cb\u597d\u5ea6<\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\">4.4 \u5176\u4ed6\u4f18\u5316<\/h3>\n\n\n\n<h3 class=\"wp-block-heading\">4.4.1 \u8bbe\u8ba1\u4f18\u5316<\/h3>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u5728 Backbone \u548c Neck\uff0c\u4f7f\u7528\u5730\u5e73\u7ebf\u5f81\u7a0b 6 \u5e73\u53f0\u9ad8\u6548\u7684 HENet \u7ed3\u6784\u4ee5\u53ca\u4f18\u5316\u540e\u7684 FPN \u7ed3\u6784\uff0c\u63d0\u9ad8\u4e86\u6a21\u578b\u5728\u7aef\u4fa7\u7684\u6027\u80fd\u8868\u73b0\uff1b<\/li>\n\n\n\n<li>\u5728\u516c\u7248 MapTR \u7684\u57fa\u7840\u4e0a\uff0c\u5f15\u5165 SD Map \u524d\u878d\u5408\uff0c\u4e3a\u5730\u56fe\u8981\u7d20\u9884\u6d4b\u63d0\u4f9b\u4e86\u4e30\u5bcc\u7684\u5148\u9a8c\u4fe1\u606f\uff0c\u6781\u5927\u63d0\u9ad8\u4e86\u6a21\u578b\u7684\u7cbe\u5ea6\u8868\u73b0\uff1b<\/li>\n\n\n\n<li>\u5728 View Transformer\uff0c\u4f7f\u7528\u6df1\u5ea6\u4f18\u5316\u8fc7\u7684 Bevformer \u66ff\u6362\u5730\u5e73\u7ebf\u652f\u6301\u4e0d\u53cb\u597d\u7684\u516c\u7248 MapTRv2 \u57fa\u4e8e LSS \u7684 BEVPoolv2 \u6765\u4f5c\u4e3a PV \u89c6\u89d2\u8f6c BEV \u89c6\u89d2\u7684\u65b9\u5f0f\uff1b<\/li>\n\n\n\n<li>\u5728 View Transformer \u7684 BEV Encoder \u6a21\u5757\u53d6\u6d88\u4e86 BEV \u7279\u5f81\u7684\u65f6\u5e8f\u878d\u5408\uff0c\u4e5f\u53d6\u6d88\u4e86 Bevformer \u65f6\u5e8f\u81ea\u6ce8\u610f\u529b\u6a21\u5757\uff0c\u6a21\u578b\u6574\u4f53\u7cbe\u5ea6\u4e0d\u4f4e\u4e8e\u516c\u7248\u57fa\u4e8e Bevformer \u7684\u7cbe\u5ea6\uff1b<\/li>\n\n\n\n<li>Head \u90e8\u5206\u53d6\u6d88\u4e86\u516c\u7248\u57fa\u4e8e\u5206\u5c42 query \u7684 decoder \u673a\u5236\uff0c\u53d6\u6d88 point query \u4f7f\u7528\u7b80\u6d01\u7684\u76f4\u63a5\u57fa\u4e8e instance query \u7684\u9884\u6d4b\uff0c\u540c\u65f6 decoder layer \u91c7\u7528\u516c\u7248 MapTRv2 \u7684\u89e3\u8026\u81ea\u6ce8\u610f\u529b\u4f18\u5316\u3002<\/li>\n<\/ol>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u603b\u7ed3\u4e0e\u5efa\u8bae<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">5.1 \u90e8\u7f72\u5efa\u8bae<\/h3>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u9075\u5faa\u786c\u4ef6\u5bf9\u9f50\u89c4\u5219\uff0c\u4e00\u822c\u7684 tensor shape \u5bf9\u9f50\u5230 2 \u7684\u5e42\u6b21\uff0cconv-like \u7684\u7b97\u5b50 H \u7ef4\u5ea6\u5bf9\u9f50\u5230 8\u3001W \u7ef4\u5ea6\u5bf9\u9f50\u5230 16\u3001C \u7ef4\u5ea6\u5bf9\u9f50\u5230 32\uff0c\u82e5\u8bbe\u8ba1\u5c3a\u5bf8\u4e0d\u6ee1\u8db3\u5bf9\u9f50\u89c4\u5219\u65f6\u4f1a\u5bf9 tensor \u81ea\u52a8\u8fdb\u884c padding\uff0c\u9020\u6210\u65e0\u6548\u7684\u7b97\u529b\u6d6a\u8d39\uff1b<\/li>\n\n\n\n<li>\u5408\u7406\u9009\u62e9 BEV Grid \u5c3a\u5bf8\uff0c\u5f81\u7a0b 6 \u5e73\u53f0\u7684\u5e26\u5bbd\u5f97\u5230\u589e\u5f3a\uff0c\u4f46\u4ecd\u9700\u8003\u8651 BEV Grid \u5c3a\u5bf8\u5bf9\u6a21\u578b\u6027\u80fd\u7684\u5f71\u54cd\uff0c\u5e76\u4e14\u7efc\u5408\u8861\u91cf\u6a21\u578b\u7cbe\u5ea6\u9884\u671f\uff0c\u9009\u62e9\u5408\u9002\u7684 BEV Grid \u5c3a\u5bf8\u4ee5\u83b7\u5f97\u6a21\u578b\u6027\u80fd\u548c\u7cbe\u5ea6\u7684\u5e73\u8861\uff1b<\/li>\n\n\n\n<li>\u4f18\u5148\u9009\u62e9\u5f81\u7a0b 6 \u5e73\u53f0\u9ad8\u6548\u7ed3\u6784\u6765\u642d\u5efa\u6a21\u578b\uff0c\u4f8b\u5982\u672c\u6587\u6240\u9009\u53d6\u7684 HENet Backbone \u548c Bevformer View Transformer\uff0c\u9ad8\u6548\u7ed3\u6784\u7ecf\u8fc7\u5728\u5f81\u7a0b 6 \u5e73\u53f0\u7684\u53cd\u590d\u4f18\u5316\u548c\u9a8c\u8bc1\uff0c\u76f8\u6bd4\u5176\u4ed6\u9009\u62e9\uff0c\u5728\u6027\u80fd\u548c\u7cbe\u5ea6\u4e0a\u53ef\u4ee5\u540c\u65f6\u53d6\u5f97\u51fa\u4f17\u7684\u6548\u679c\u3002<\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\">5.2 \u603b\u7ed3<\/h3>\n\n\n\n<p>\u672c\u6587\u901a\u8fc7\u5bf9 MapTR \u8fdb\u884c\u5730\u5e73\u7ebf\u91cf\u5316\u90e8\u7f72\u7684\u4f18\u5316\uff0c\u4f7f\u5f97\u6a21\u578b\u5728\u5f81\u7a0b 6 \u8ba1\u7b97\u5e73\u53f0\u4e0a\u7528\u8f83\u4f4e\u7684\u91cf\u5316\u7cbe\u5ea6\u635f\u5931\uff0c\u6700\u4f18\u83b7\u5f97\u5f81\u7a0b 6M \u5355\u6838 93.77 FPS \u7684\u90e8\u7f72\u6027\u80fd\u3002\u540c\u65f6\uff0cMapTR \u7cfb\u5217\u7684\u90e8\u7f72\u7ecf\u9a8c\u53ef\u4ee5\u63a8\u5e7f\u5230\u5176\u4ed6\u76f8\u4f3c\u7ed3\u6784\u6216\u76f8\u4f3c\u4f7f\u7528\u573a\u666f\u6a21\u578b\u7684\u90e8\u7f72\u4e2d\u3002<\/p>\n\n\n\n<p>\u5bf9\u4e8e\u5730\u5e73\u7ebf MapTR \u53c2\u8003\u7b97\u6cd5\u6a21\u578b\uff0c\u7ed3\u5408 Sparse Bev \u7b49\u7684\u4f18\u5316\u65b9\u5411\u4ecd\u5728\u63a2\u7d22\u548c\u5b9e\u8df5\u4e2d\uff0cStay Tuned\uff01<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001\u9644\u5f55<\/h2>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u516c\u7248\u8bba\u6587\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/arxiv.org\/abs\/2208.14437\" target=\"_blank\" rel=\"noreferrer noopener\">MapTR<\/a>\uff1b<\/li>\n\n\n\n<li>\u516c\u7248\u6a21\u578b\u6e90\u7801\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/hustvl\/MapTR\" target=\"_blank\" rel=\"noreferrer noopener\">GitHub-MapTR<\/a>\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u539f\u6587\u94fe\u63a5\uff1ahttps:\/\/zhuanlan.z [&hellip;]<\/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-30436","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\/30436"}],"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=30436"}],"version-history":[{"count":1,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/30436\/revisions"}],"predecessor-version":[{"id":30437,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/30436\/revisions\/30437"}],"wp:attachment":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/media?parent=30436"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/categories?post=30436"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/tags?post=30436"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}