{"id":5489,"date":"2026-09-20T23:11:50","date_gmt":"2026-09-20T15:11:50","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=5489"},"modified":"2026-09-20T23:11:50","modified_gmt":"2026-09-20T15:11:50","slug":"202609-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-deepasa-an-object-oriented-multi-purpose-network-for-audittory-scene-analysis","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202609-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-deepasa-an-object-oriented-multi-purpose-network-for-audittory-scene-analysis\/","title":{"rendered":"202609 \u8bba\u6587\u7814\u8bfb-DeepASA: An Object-Oriented Multi-Purpose Network for Audittory Scene Analysis"},"content":{"rendered":"\n

Author: <\/em><\/strong>Dongheon Lee, Younghoo Kwon, Jung-Woo Choi<\/em>
Unit:Korea Advanced Institute of Science and Technology\uff08KAIST\uff09<\/em><\/strong>
Journal: <\/em><\/strong>NeurIPS<\/em>
Year: <\/em><\/strong>2025<\/em><\/p>\n\n\n\n

\u4e00\u3001\u7814\u7a76\u80cc\u666f<\/h2>\n\n\n\n
\"\"<\/figure>\n\n\n\n

\u4e8c\u3001\u6838\u5fc3\u8d21\u732e<\/h2>\n\n\n\n
\"\"<\/figure>\n\n\n\n

\u4e09\u3001\u65b9\u6cd5<\/h2>\n\n\n\n

(1)\u5bf9\u8f93\u5165\u7684\u591a\u901a\u9053\u6df7\u5408\u97f3\u9891\uff0c\u9996\u5148\u4f7f\u7528 Dynamic STFT \u63d0\u53d6\u590d\u6570\u65f6\u9891\u7279\u5f81\u3002\u6a21\u578b\u9010\u5e27\u9884\u6d4b\u9ad8\u65af\u7a97\u7684\u4f4d\u7f6e\u548c\u5bbd\u5ea6\uff0c\u4f7f\u5206\u6790\u7a97\u53e3\u80fd\u591f\u6839\u636e\u58f0\u97f3\u5185\u5bb9\u52a8\u6001\u8c03\u6574\u65f6\u95f4\u548c\u9891\u7387\u5206\u8fa8\u80fd\u529b\u3002<\/p>\n\n\n\n

(2)\u968f\u540e\u5229\u7528\u4fee\u6539\u540e\u7684 DeFT-Mamba \u805a\u5408\u65f6\u95f4\u3001\u9891\u7387\u548c\u591a\u901a\u9053\u7a7a\u95f4\u5173\u7cfb\uff0c\u518d\u901a\u8fc7 Object Splitter \u5c06\u6574\u4e2a\u58f0\u573a\u8868\u793a\u62c6\u5206\u4e3a\u591a\u4e2a foreground object \u548c\u4e00\u4e2a\u72ec\u7acb\u7684 noise object\u3002<\/p>\n\n\n\n

(3)\u6bcf\u4e2a\u58f0\u97f3\u5bf9\u8c61\u5206\u522b\u8fdb\u5165 Audio\u3001SED \u548c DoA \u89e3\u7801\u5668\uff0c\u9884\u6d4b\u5bf9\u5e94\u7684 direct\/reverb \u6ce2\u5f62\u3001\u58f0\u97f3\u7c7b\u522b\u3001\u6d3b\u52a8\u72b6\u6001\u548c\u7a7a\u95f4\u65b9\u5411\uff1b\u7b2c\u4e00\u8f6e\u7ed3\u679c\u518d\u901a\u8fc7 CoI \u548c TCM \u76f8\u4e92\u6821\u9a8c\uff0c\u5e76\u53cd\u9988\u5230 object feature \u4e2d\u8fdb\u884c\u7b2c\u4e8c\u6b21\u63a8\u7406<\/p>\n\n\n\n

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

\u56db\u3001\u5b9e\u9a8c<\/h2>\n\n\n\n

1.\u5b9e\u9a8c\u51c6\u5907<\/p>\n\n\n\n

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

2.\u5b9e\u9a8c\u7ed3\u679c\u2014\u2014Ablation Study of DeepASA<\/strong><\/strong><\/strong><\/p>\n\n\n\n

(1)\u968f\u7740 SED\/DoA decoder\u3001Noise decoder\u3001Direct\/Reverb decoder\u3001Dynamic STFT \u548c CoI \u4f9d\u6b21\u52a0\u5165\uff0c\u6a21\u578b\u4ece\u5355\u4e00\u58f0\u97f3\u5206\u79bb\u9010\u6e10\u5f62\u6210\u5b8c\u6574\u7684 object-oriented auditory scene analysis \u80fd\u529b\u3002<\/p>\n\n\n\n

(2)\u663e\u5f0f\u5efa\u7acb Noise object \u80fd\u591f\u5e2e\u52a9\u6a21\u578b\u533a\u5206\u524d\u666f\u58f0\u97f3\u548c\u80cc\u666f\u566a\u58f0\uff1bDirect\/Reverb \u5206\u5f00\u5efa\u6a21\u5219\u6709\u52a9\u4e8e\u524a\u5f31\u6df7\u54cd\u5bf9\u7a7a\u95f4\u5b9a\u4f4d\u7684\u5f71\u54cd\u3002<\/p>\n\n\n\n

(3)Dynamic STFT \u6539\u5584\u4e0d\u540c\u65f6\u95f4\u5c3a\u5ea6\u58f0\u97f3\u7684\u8868\u793a\uff0c\u800c CoI \u8fdb\u4e00\u6b65\u5229\u7528 SED \u4e0e DoA \u7684\u4e92\u8865\u5173\u7cfb\u4fee\u6b63 object feature\uff0c\u6700\u7ec8\u8fbe\u5230 SI-SDRi 11.2 dB\u3001SELD score 0.206\u3002<\/p>\n\n\n\n

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

3.\u5b9e\u9a8c\u7ed3\u679c\u2014\u2014Generalization to Spatial Audio Benchmarks<\/strong><\/strong><\/p>\n\n\n\n

(1)\u5728 MC-FUSS \u901a\u7528\u58f0\u97f3\u5206\u79bb\u4efb\u52a1\u4e0a\uff0cASA2 \u9884\u8bad\u7ec3\u540e\u7684 DeepASA \u8fbe\u5230 18.5 dB SI-SDRi\uff0c\u5e76\u4e14\u5728\u591a\u58f0\u6e90\u91cd\u53e0\u573a\u666f\u4e2d\u4ecd\u4fdd\u6301\u8f83\u597d\u7684\u5206\u79bb\u80fd\u529b\uff0c\u8bf4\u660e object representation \u53ef\u4ee5\u8fc1\u79fb\u5230\u5176\u4ed6\u58f0\u97f3\u5206\u79bb\u6570\u636e\u96c6\u3002<\/p>\n\n\n\n

(2)\u5728 STARSS23 SELD \u4efb\u52a1\u4e0a\uff0cDeepASA + CoI \u83b7\u5f97 0.253 \u7684\u7efc\u5408 SELD score \u548c 9.8\u00b0 \u7684\u5b9a\u4f4d\u8bef\u5dee\uff0c\u8bf4\u660e\u8be5\u65b9\u6cd5\u4e0d\u4ec5\u9002\u7528\u4e8e\u4f5c\u8005\u6784\u5efa\u7684 ASA2\uff0c\u4e5f\u80fd\u591f\u8fc1\u79fb\u5230\u771f\u5b9e\u7a7a\u95f4\u97f3\u9891\u6570\u636e\u3002<\/p>\n\n\n\n

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

\u4e94\u3001\u603b\u7ed3<\/h2>\n\n\n\n

(1)\u63d0\u51fa DeepASA \u7edf\u4e00\u542c\u89c9\u573a\u666f\u5206\u6790\u6846\u67b6\uff0c\u5c06\u58f0\u97f3\u5206\u79bb\u3001\u53bb\u6df7\u54cd\u3001\u566a\u58f0\u4f30\u8ba1\u3001\u58f0\u97f3\u4e8b\u4ef6\u68c0\u6d4b\u3001\u5206\u7c7b\u548c\u7a7a\u95f4\u5b9a\u4f4d\u6574\u5408\u5230\u540c\u4e00\u4e2a\u6a21\u578b\u4e2d\u3002<\/p>\n\n\n\n

(2)\u63d0\u51fa Object-Oriented Processing\uff0c\u4f7f\u58f0\u97f3\u7684\u6ce2\u5f62\u3001\u7c7b\u522b\u3001\u6d3b\u52a8\u72b6\u6001\u548c\u7a7a\u95f4\u4f4d\u7f6e\u56f4\u7ed5\u540c\u4e00\u4e2a object \u8868\u793a\u8fdb\u884c\u7ec4\u7ec7\uff0c\u51cf\u5c11\u4f20\u7edf track-wise processing \u4e2d\u7684\u4fe1\u606f\u5173\u8054\u95ee\u9898\u3002<\/p>\n\n\n\n

(3)\u63d0\u51fa Chain-of-Inference\uff0c\u901a\u8fc7 SED\u3001DoA \u7b49\u4e0d\u540c\u542c\u89c9\u7ebf\u7d22\u4e4b\u95f4\u7684\u4ea4\u4e92\uff0c\u5bf9\u7b2c\u4e00\u6b21\u5f62\u6210\u7684\u4e0d\u5b8c\u6574 object representation \u8fdb\u884c\u8fdb\u4e00\u6b65\u4fee\u6b63\u3002<\/p>\n\n\n\n

(4)DeepASA \u7684\u6838\u5fc3\u610f\u4e49\u5e76\u4e0d\u53ea\u662f\u201c\u540c\u65f6\u5b8c\u6210\u591a\u4e2a\u4efb\u52a1\u201d\uff0c\u800c\u662f\u5c06\u542c\u89c9\u573a\u666f\u4ece\u591a\u4e2a\u72ec\u7acb\u9884\u6d4b\u7ed3\u679c\u8fdb\u4e00\u6b65\u7ec4\u7ec7\u4e3a\u201c\u58f0\u97f3\u5bf9\u8c61 + \u5bf9\u8c61\u5c5e\u6027 + \u5c5e\u6027\u5173\u7cfb\u201d\u7684\u7ed3\u6784\u5316\u8868\u793a\u3002<\/p>\n\n\n\n

\u516d\u3001\u5bf9\u9f50\u601d\u8003<\/h2>\n\n\n\n

1.\u65b9\u6cd5\u521b\u65b0\u70b9\u2014\u2014\u5168\u57df\u6570\u636e\u62d3\u6251\u5efa\u6a21: DeepASA \u4e0d\u518d\u5c06\u8bed\u4e49\u3001\u7a7a\u95f4\u548c\u58f0\u97f3\u5185\u5bb9\u4f5c\u4e3a\u4e92\u4e0d\u76f8\u5173\u7684\u8f93\u51fa\uff0c\u800c\u662f\u56f4\u7ed5\u5355\u4e2a\u58f0\u6e90\u5efa\u7acb\u201cObject\u2014Attribute\u2014Relation\u201d\u7ed3\u6784\uff0c\u4e3a\u8fdb\u4e00\u6b65\u6784\u5efa\u58f0\u6e90\u7ea7\u7ed3\u6784\u5316\u58f0\u573a\u8868\u793a\u63d0\u4f9b\u4e86\u53c2\u8003\u3002<\/p>\n\n\n\n

2.\u6280\u672f\u76ee\u6807\u70b9\u2014\u2014\u865a\u5b9e\u7a7a\u95f4\u5b6a\u751f\u63a8\u6f14: \u6a21\u578b\u80fd\u591f\u540c\u65f6\u63cf\u8ff0\u201c\u662f\u4ec0\u4e48\u58f0\u97f3\u3001\u4ec0\u4e48\u65f6\u5019\u5b58\u5728\u3001\u6765\u81ea\u54ea\u91cc\u3001\u5bf9\u5e94\u4ec0\u4e48\u58f0\u97f3\u4fe1\u53f7\u201d\uff0c\u76f8\u6bd4\u5355\u4e00\u5168\u5c40 embedding\uff0c\u66f4\u63a5\u8fd1\u53ef\u89e3\u91ca\u7684\u58f0\u573a\u5bf9\u8c61\u5efa\u6a21\u3002<\/p>\n\n\n\n

3.\u4e2d\u8bd5\u4ea7\u54c1\u70b9\u2014\u2014\u57c3\u89c5\u6587\u65c5: \u53ef\u5c06\u590d\u6742\u73af\u5883\u58f0\u97f3\u89e3\u6790\u4e3a\u591a\u4e2a\u5177\u6709\u7c7b\u522b\u3001\u65f6\u95f4\u548c\u7a7a\u95f4\u4fe1\u606f\u7684\u58f0\u6e90\u5bf9\u8c61\uff0c\u4e3a\u666f\u533a\u58f0\u97f3\u8bc6\u522b\u3001\u58f0\u6e90\u5b9a\u4f4d\u53ca\u58f0\u666f\u8bc4\u4ef7\u63d0\u4f9b\u7ed3\u6784\u5316\u8f93\u5165\uff0c\u5e76\u8fdb\u4e00\u6b65\u89e3\u91ca\u201c\u54ea\u4e9b\u58f0\u97f3\u3001\u4ece\u54ea\u91cc\u51fa\u73b0\u3001\u6301\u7eed\u591a\u4e45\u201d\u5f71\u54cd\u5f53\u524d\u58f0\u666f\u4f53\u9a8c\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

Author: Dongheon Lee, Younghoo Kwon, Jung-Woo ChoiUnit: […]<\/p>\n","protected":false},"author":74,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-5489","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/5489","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/users\/74"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=5489"}],"version-history":[{"count":1,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/5489\/revisions"}],"predecessor-version":[{"id":5505,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/5489\/revisions\/5505"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=5489"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=5489"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=5489"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}