{"id":13390,"date":"2026-10-15T21:21:50","date_gmt":"2026-10-15T13:21:50","guid":{"rendered":"https:\/\/infernews.com\/blog\/urbex\/"},"modified":"2026-10-15T21:24:12","modified_gmt":"2026-10-15T13:24:12","slug":"urbex","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/urbex\/","title":{"rendered":"Urbex \u8b93\u57ce\u5e02\u7d1a 3D \u5b9a\u4f4d\u6539\u7528\u4e3b\u52d5\u641c\u5c0b"},"content":{"rendered":"\n
\"Urbex<\/figure>\n\n\n\n

\u9762\u5c0d\u57ce\u5e02\u7d1a 3D \u5834\u666f\uff0c\u9010\u4e00\u6bd4\u8f03\u9810\u5148\u7522\u751f\u7684\u7269\u4ef6\u5019\u9078\u6846\u5bb9\u6613\u62d6\u6162\u641c\u5c0b\uff0c\u4e5f\u6703\u53d7\u5019\u9078\u54c1\u8cea\u9650\u5236\u3002Urbex \u5c6c\u65bc\u958b\u6e90\u7684 Agentic \u8996\u89ba\u6a21\u578b\u9805\u76ee\uff0c\u5c07 3D grounding \u6539\u6210\u4e3b\u52d5\u7a7a\u9593\u641c\u5c0b\uff0c\u8b93 vision-language agent \u6309\u9700\u8981\u67e5\u8a62\u5730\u6a19\u3001\u7e2e\u653e\u5340\u57df\u3001\u5207\u63db\u659c\u8996\u89d2\uff0c\u518d\u63d0\u4ea4\u76ee\u6a19 bounding box\u3002<\/p>\n\n\n\n

\u5b83\u4e0d\u662f\u53ea\u9760\u4e00\u6b21\u5f71\u50cf\u5224\u65b7\uff0c\u800c\u662f\u5728\u4e92\u52d5\u5f0f multi-view environment \u9010\u6b65\u6536\u96c6\u8b49\u64da\u3002Reinforcement learning \u540c\u6642\u8a13\u7df4\u5b9a\u4f4d\u548c\u641c\u5c0b\u6c7a\u7b56\uff0c\u6a21\u578b\u6703\u5b78\u7fd2\u61c9\u8a72\u6aa2\u67e5\u54ea\u88cf\u3001\u4f55\u6642\u7e2e\u7a84\u7bc4\u570d\uff0c\u4ee5\u53ca\u5e7e\u6642\u8db3\u4ee5\u4f5c\u7b54\uff1b\u9019\u514d\u9664\u5c0d precomputed instance proposals \u548c ground-truth boxes \u7684\u4f9d\u8cf4\uff0c\u4f46\u6bcf\u6b21\u641c\u5c0b\u4ecd\u8981\u627f\u53d7\u591a\u6b65\u64cd\u4f5c\u6210\u672c\u3002<\/p>\n\n\n<\/div>\n\n\n

CityRefer \u6e2c\u8a66\u4e2d\uff0cUrbex \u7684 ND Hit@1\u3001Hit@3\u3001Hit@5 \u70ba 25.8\u300139.2\u300142.3\uff0cNO \u5c0d\u61c9\u70ba 23.9\u300136.9\u300140.7\uff0c\u641c\u5c0b\u6642\u9593\u7d04 3.2 \u79d2\u3002\u76f8\u8f03 CityAnchor \u7684 32.1 \u79d2\uff0c\u901f\u5ea6\u660e\u986f\u8f03\u5feb\uff1b\u6a21\u578b\u4ea6\u5728 STPLS3D-Refer \u53d6\u5f97 zero-shot \u7d50\u679c\uff0c\u53cd\u6620\u8de8\u5834\u666f\u6cdb\u5316\u6f5b\u529b\uff0c\u4f46\u4e0d\u80fd\u76f4\u63a5\u7b49\u540c\u65bc\u6240\u6709\u57ce\u5e02\u74b0\u5883\u90fd\u540c\u6a23\u7a69\u5b9a\u3002<\/p>\n\n\n\n

\u2022 \u652f\u63f4 CityRefer \u8a13\u7df4\u53ca\u55ae\u67e5\u8a62\u4e92\u52d5\u5f0f 3D \u63a8\u7406
\n\u2022 \u63d0\u4f9b Hugging Face \u7684 Urbex-4B \u6a21\u578b\u548c Urbex-Data \u6578\u64da\u96c6
\n\u2022 \u53ef\u900f\u904e Colab notebook \u8a66\u7528\u9810\u6e2c\u8996\u5716
\n\u2022 \u5b89\u88dd\u9700\u8981 Linux\u3001Python 3.11\u3001CUDA 12.8 \u53ca C++ \u7de8\u8b6f\u5668<\/p>\n\n\n\n

\u7814\u7a76\u57ce\u5e02\u5efa\u6a21\u30013D \u8996\u89ba\u5b9a\u4f4d\u6216\u591a\u6a21\u614b\u4ee3\u7406\u7684\u5718\u968a\uff0c\u8f03\u9069\u5408\u7528\u5b83\u6e2c\u8a66 candidate-free grounding\uff1b\u60f3\u5feb\u901f\u9a57\u8b49\u6982\u5ff5\u7684\u4eba\u5247\u53ef\u5148\u7528 Colab\u3002\u5b8c\u6574\u8a13\u7df4\u9700\u8981 CUDA 12.8\u3001Megatron Core\u3001Transformer Engine \u548c FlashAttention\uff0c\u786c\u4ef6\u53ca\u74b0\u5883\u9580\u6abb\u660e\u986f\u9ad8\u65bc\u55ae\u7d14\u63a8\u7406\uff0c\u56e0\u6b64\u672c\u5730\u90e8\u7f72\u524d\u61c9\u5148\u78ba\u8a8d GPU\u3001\u7de8\u8b6f\u5de5\u5177\u548c\u6578\u64da\u4e0b\u8f09\u6d41\u7a0b\u3002<\/p>\n\n\n\n

\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 GitHub<\/strong><\/a> \u00b7 \u6a21\u578b<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

Urbex \u8b93\u8996\u89ba\u8a9e\u8a00\u4ee3\u7406\u5728\u57ce\u5e02\u7d1a 3D \u5834\u666f\u81ea\u884c\u641c\u5c0b\u8b49\u64da\uff0c\u6bcb\u9808\u9810\u5148\u5efa\u7acb\u7269\u4ef6\u5019\u9078\u6846\uff0c\u5b9a\u4f4d\u80fd\u529b\u8207\u6548\u7387\u53d6\u6368\u66f4\u503c\u5f97\u6aa2\u8996\u3002<\/p>\n","protected":false},"author":8,"featured_media":13389,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","wpai_meta_description":"","footnotes":""},"categories":[133,171,116,76,149,119,127,185,179,30,38,132,168,152,199,214],"tags":[],"class_list":["post-13390","post","type-post","status-publish","format-standard","hentry","category-133","category-171","category-agentic","category-76","category-149","category-119","category-127","category-qwen","category-nvidia","category-image","category-38","category-3d","category-linux","category-python","category-dataset-","category-reinforcement-learning"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13390","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=13390"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13390\/revisions"}],"predecessor-version":[{"id":13391,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13390\/revisions\/13391"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/13389"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=13390"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=13390"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=13390"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}