{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"Faster R-CNN：让候选区域也成为预测","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html","markdown":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.md","pages":[{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","unit":"类别与位置怎样成为可学习的区域目标？","unitId":"learnable-box-prediction","chapter":"II","number":1},{"id":"anchor-training-targets","title":"给正例、背景和忽略项分配任务","unit":"类别与位置怎样成为可学习的区域目标？","unitId":"learnable-box-prediction","chapter":"II","number":2},{"id":"detector-lab10-evidence","title":"进入 LAB10：先读失败图，再解释高分","unit":"检测分数很高，错误究竟在哪里？","unitId":"detection-errors-and-evidence","chapter":"II","number":3},{"id":"dense-detector-shared-problems","title":"一阶段检测简化哪一段，保留哪些问题？","unit":"选读：输出组织与训练分配还可以怎样改变？","unitId":"spatial-training-extensions","chapter":"II","number":4}],"aliases":[{"oldId":"explanation","targetId":"reference-proposal-detection","operation":"split-explanation-range"},{"oldId":"b-186519f1b6f69e3db2ee6247","targetId":"b-186519f1b6f69e3db2ee6247","page":"reference-proposal-detection","operation":"move","beforeSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","afterSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","reason":"Same source body in authored semantic page"},{"oldId":"b-20a30a24e745c5c7804b77bb","targetId":"b-20a30a24e745c5c7804b77bb","page":"reference-proposal-detection","operation":"move","beforeSHA256":"fa5fee02eec78a0fc68677365db555ff86cf68a2947a56e2a2f54c2a789e96f2","afterSHA256":"fa5fee02eec78a0fc68677365db555ff86cf68a2947a56e2a2f54c2a789e96f2","reason":"Same source body in authored semantic page"},{"oldId":"b-4b845ae67f6bca64d7a9262d","targetId":"b-4b845ae67f6bca64d7a9262d","page":"reference-proposal-detection","operation":"move","beforeSHA256":"6b45cf47a23e9eec2f15663b9a1624db57adbc4f2a6825dd6bc21c96bfb93aa4","afterSHA256":"6b45cf47a23e9eec2f15663b9a1624db57adbc4f2a6825dd6bc21c96bfb93aa4","reason":"Same source body in authored semantic page"},{"oldId":"b-b41f25904357eb5693af08bf","targetId":"b-b41f25904357eb5693af08bf","page":"reference-proposal-detection","operation":"rewrite","beforeSHA256":"cdf793d7f3ca501e452511b276ca0587b1f1057e84bc699d01db574c5eca790b","afterSHA256":"682d9ad365a9a0ea524a22340299a5f3812904ef83a368c0161ca3325f532cec","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-12220d5b859b781cdec0e1b7","targetId":"b-12220d5b859b781cdec0e1b7","page":"reference-proposal-detection","operation":"move","beforeSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","afterSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","reason":"Same source body in authored semantic page"},{"oldId":"b-dabd833266f6e275ec84f502","targetId":"b-dabd833266f6e275ec84f502","page":"reference-proposal-detection","operation":"move","beforeSHA256":"febdeb92cc5c7d107b2125830e7f25e5103c9845e4447dd8ccda5f273e85c59b","afterSHA256":"febdeb92cc5c7d107b2125830e7f25e5103c9845e4447dd8ccda5f273e85c59b","reason":"Same source body in authored semantic page"},{"oldId":"b-bf6003d8695683a668235d24","targetId":"b-bf6003d8695683a668235d24","page":"reference-proposal-detection","operation":"move","beforeSHA256":"8d8642c37d0a16aa2c09a67fd19a693edc3211397fb9026649867e38a51c165b","afterSHA256":"8d8642c37d0a16aa2c09a67fd19a693edc3211397fb9026649867e38a51c165b","reason":"Same source body in authored semantic page"},{"oldId":"b-bfb211a278f543d28f330665","targetId":"b-bfb211a278f543d28f330665","page":"reference-proposal-detection","operation":"move","beforeSHA256":"ff6018e332450f304b746defc546d270e531208c2bd0354e7e32a269661363c1","afterSHA256":"ff6018e332450f304b746defc546d270e531208c2bd0354e7e32a269661363c1","reason":"Same source body in authored semantic page"},{"oldId":"b-9700ffd63e3598027017fc2c","targetId":"b-9700ffd63e3598027017fc2c","page":"reference-proposal-detection","operation":"move","beforeSHA256":"ec5bbb4f56f3be17858a079f938ad06774395a07f9f46cda16ebf9c2538d0144","afterSHA256":"ec5bbb4f56f3be17858a079f938ad06774395a07f9f46cda16ebf9c2538d0144","reason":"Same source body in authored semantic page"},{"oldId":"b-3d61f626d5daf110bda0c610","targetId":"b-3d61f626d5daf110bda0c610","page":"reference-proposal-detection","operation":"move","beforeSHA256":"87e8f6e090d6b2bb0be5d6005c6ef7fae57d2b05e1c4df5c825928472ed755bf","afterSHA256":"87e8f6e090d6b2bb0be5d6005c6ef7fae57d2b05e1c4df5c825928472ed755bf","reason":"Same source body in authored semantic page"},{"oldId":"b-d05cd9d1f66d10cb8526492f","targetId":"b-d05cd9d1f66d10cb8526492f","page":"reference-proposal-detection","operation":"move","beforeSHA256":"d39d8ade1a97f74de22f5548d14b4a3ca83a4cff593b2c987f0b8ead3d6a7d6e","afterSHA256":"d39d8ade1a97f74de22f5548d14b4a3ca83a4cff593b2c987f0b8ead3d6a7d6e","reason":"Same source body in authored semantic page"},{"oldId":"b-171e4ea2e1cf37b2ee0113ea","targetId":"b-171e4ea2e1cf37b2ee0113ea","page":"reference-proposal-detection","operation":"move","beforeSHA256":"bbfc204ffba300b74c47d93813472b503913a6c38c28d2a201fabd8502b3bb98","afterSHA256":"bbfc204ffba300b74c47d93813472b503913a6c38c28d2a201fabd8502b3bb98","reason":"Same source body in authored semantic page"},{"oldId":"b-3a30bc257b48623626999482","targetId":"b-3a30bc257b48623626999482","page":"reference-proposal-detection","operation":"move","beforeSHA256":"9907f492b45d33d2cd095cf93e4f2992b615967bc013d051ce64b24f87065754","afterSHA256":"9907f492b45d33d2cd095cf93e4f2992b615967bc013d051ce64b24f87065754","reason":"Same source body in authored semantic page"},{"oldId":"b-0158a60b62fbc2e82f98bb50","targetId":"b-0158a60b62fbc2e82f98bb50","page":"reference-proposal-detection","operation":"move","beforeSHA256":"b509fe9b9654e59b2393a987864d15730dba547c20e00743ea434b4dd911fbc9","afterSHA256":"b509fe9b9654e59b2393a987864d15730dba547c20e00743ea434b4dd911fbc9","reason":"Same source body in authored semantic page"},{"oldId":"b-2bfac8a38ce5d482cef7c0e9","targetId":"b-2bfac8a38ce5d482cef7c0e9","page":"reference-proposal-detection","operation":"move","beforeSHA256":"19ab452623b51a31d2b44d4c7e9d38c8d7fd345a5d0d975abe5a42661bea965d","afterSHA256":"19ab452623b51a31d2b44d4c7e9d38c8d7fd345a5d0d975abe5a42661bea965d","reason":"Same source body in authored semantic page"},{"oldId":"b-16a7a5fda344ac718152903b","targetId":"b-16a7a5fda344ac718152903b","page":"reference-proposal-detection","operation":"move","beforeSHA256":"84fada740d51b378cea3e0e47e6651c3bb1828c75c12514430355bdb75da67d4","afterSHA256":"84fada740d51b378cea3e0e47e6651c3bb1828c75c12514430355bdb75da67d4","reason":"Same source body in authored semantic page"},{"oldId":"b-2c8e8ec10d779a7959ac5235","targetId":"b-2c8e8ec10d779a7959ac5235","page":"reference-proposal-detection","operation":"move","beforeSHA256":"b7fb3753c687fe3752b886d2f752f2ab90796048541029814935ab403f4031e4","afterSHA256":"b7fb3753c687fe3752b886d2f752f2ab90796048541029814935ab403f4031e4","reason":"Same source body in authored semantic page"},{"oldId":"b-71ecab44188d91b44d011801","targetId":"b-71ecab44188d91b44d011801","page":"reference-proposal-detection","operation":"move","beforeSHA256":"91124d0e21521d83268cc3ffa1876dac09f5748aa130e072ee53b1a27cb9e450","afterSHA256":"91124d0e21521d83268cc3ffa1876dac09f5748aa130e072ee53b1a27cb9e450","reason":"Same source body in authored semantic page"},{"oldId":"b-d1dd5dac3ea6e2fb9d4f236e","targetId":"b-d1dd5dac3ea6e2fb9d4f236e","page":"reference-proposal-detection","operation":"move","beforeSHA256":"5a6614c9865a2a725016da4bdb0f7756f578db6952ab280a0accbeb12cb5c7c8","afterSHA256":"5a6614c9865a2a725016da4bdb0f7756f578db6952ab280a0accbeb12cb5c7c8","reason":"Same source body in authored semantic page"},{"oldId":"b-c5b9177a9a22aeb431cb2791","targetId":"b-c5b9177a9a22aeb431cb2791","page":"reference-proposal-detection","operation":"move","beforeSHA256":"4cf00996a0f6282ac8d4f2c1ef4f559764812bb995da338db000ec5a7e6c1330","afterSHA256":"4cf00996a0f6282ac8d4f2c1ef4f559764812bb995da338db000ec5a7e6c1330","reason":"Same source body in authored semantic page"},{"oldId":"b-acfe7f50e1e8b461a2a75a59","targetId":"b-acfe7f50e1e8b461a2a75a59","page":"reference-proposal-detection","operation":"move","beforeSHA256":"b464b01fa76793dd871c621138c876cd92257e297f945300d9dbbd5206e92eac","afterSHA256":"b464b01fa76793dd871c621138c876cd92257e297f945300d9dbbd5206e92eac","reason":"Same source body in authored semantic page"},{"oldId":"b-9bad1749ccd77ae4fb1e627c","targetId":"b-9bad1749ccd77ae4fb1e627c","page":"anchor-training-targets","operation":"move","beforeSHA256":"7af001429caa370b259fa0b7820036e8a5a1052d83c445a1e2ab6fbbda8c7c93","afterSHA256":"7af001429caa370b259fa0b7820036e8a5a1052d83c445a1e2ab6fbbda8c7c93","reason":"Same source body in authored semantic page"},{"oldId":"b-94169f26c39be5ca7ebae497","targetId":"b-94169f26c39be5ca7ebae497","page":"anchor-training-targets","operation":"move","beforeSHA256":"290dba49c527e68e2aba4a2ee0b4d28e626d8e489edab4605ea5a235e8ea7fb2","afterSHA256":"290dba49c527e68e2aba4a2ee0b4d28e626d8e489edab4605ea5a235e8ea7fb2","reason":"Same source body in authored semantic page"},{"oldId":"b-d80f2878b651375aec18736e","targetId":"b-d80f2878b651375aec18736e","page":"anchor-training-targets","operation":"move","beforeSHA256":"9950be4406a476d8d766763df53ea513bbe7eed02f34aff4f837c30a3d045b5c","afterSHA256":"9950be4406a476d8d766763df53ea513bbe7eed02f34aff4f837c30a3d045b5c","reason":"Same source body in authored semantic page"},{"oldId":"b-294c9d12e24bab16c72b8945","targetId":"b-294c9d12e24bab16c72b8945","page":"anchor-training-targets","operation":"move","beforeSHA256":"157cb5eec4f8576589386132685fedffe3cc7d7ba664168060367d6e5a58232b","afterSHA256":"157cb5eec4f8576589386132685fedffe3cc7d7ba664168060367d6e5a58232b","reason":"Same source body in authored semantic page"},{"oldId":"b-522412585d246c7a10026eb2","targetId":"b-522412585d246c7a10026eb2","page":"anchor-training-targets","operation":"move","beforeSHA256":"767e9cdba079c3d5360d01ca6637ff47b625f805ea988b0a945b3600317d283c","afterSHA256":"767e9cdba079c3d5360d01ca6637ff47b625f805ea988b0a945b3600317d283c","reason":"Same source body in authored semantic page"},{"oldId":"b-ecc4865a85ab9a1ade3eb319","targetId":"b-ecc4865a85ab9a1ade3eb319","page":"detector-lab10-evidence","operation":"move","beforeSHA256":"76f6a9ca0e69cd8094fa23bd7ccbf42fb6bb705fb6fa0719bfe6760e58a0769f","afterSHA256":"76f6a9ca0e69cd8094fa23bd7ccbf42fb6bb705fb6fa0719bfe6760e58a0769f","reason":"Same source body in authored semantic page"},{"oldId":"b-cb691a0fc8abcef8054b126e","targetId":"b-cb691a0fc8abcef8054b126e","page":"detector-lab10-evidence","operation":"move","beforeSHA256":"2721b637cb6f94558f6713c359032adafb78e252bc13f52e76ce70874291e64b","afterSHA256":"2721b637cb6f94558f6713c359032adafb78e252bc13f52e76ce70874291e64b","reason":"Same source body in authored semantic page"},{"oldId":"b-447067e18fcf166999ea4137","targetId":"b-447067e18fcf166999ea4137","page":"detector-lab10-evidence","operation":"move","beforeSHA256":"7f64f21b960df2c59e03e9c18028330d1e9a641549b00585d29c5ca6e0f9b5b3","afterSHA256":"7f64f21b960df2c59e03e9c18028330d1e9a641549b00585d29c5ca6e0f9b5b3","reason":"Same source body in authored semantic page"},{"oldId":"b-5b30369ff133e6b8dabd80a3","targetId":"b-5b30369ff133e6b8dabd80a3","page":"detector-lab10-evidence","operation":"rewrite","beforeSHA256":"0ca8335cec01613a77310772ede6cd2201e74b27c01699f9c6c5125faa8c5786","afterSHA256":"ea3f2b25b776460f4e0e045b665251387ea0450a36a4ff6475b9f0188e257153","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-dda7233a52b87fa2bbcd5d55","targetId":"b-dda7233a52b87fa2bbcd5d55","page":"detector-lab10-evidence","operation":"move","beforeSHA256":"8205b30be23a99f3c1540701846c79ab37442feeb48cf7193d02963131b7f9ab","afterSHA256":"8205b30be23a99f3c1540701846c79ab37442feeb48cf7193d02963131b7f9ab","reason":"Same source body in authored semantic page"},{"oldId":"b-b922ccb95bec0009002c3d4b","targetId":"b-b922ccb95bec0009002c3d4b","page":"detector-lab10-evidence","operation":"move","beforeSHA256":"1caaa88a4dc7420a22d2566dc2b1a3e78a78af9d5f019befb84d7a50e0d3a296","afterSHA256":"1caaa88a4dc7420a22d2566dc2b1a3e78a78af9d5f019befb84d7a50e0d3a296","reason":"Same source body in authored semantic page"},{"oldId":"b-5aeffabcf350b1df8c2f3e65","targetId":"b-5aeffabcf350b1df8c2f3e65","page":"dense-detector-shared-problems","operation":"move","beforeSHA256":"6d7a9f0134bacd810263ef2e018b0b9caefa60be8a73e1c4d75a595f843edc73","afterSHA256":"6d7a9f0134bacd810263ef2e018b0b9caefa60be8a73e1c4d75a595f843edc73","reason":"Same source body in authored semantic page"},{"oldId":"b-5e05bacb7c023a91ee20a16a","targetId":"b-5e05bacb7c023a91ee20a16a","page":"dense-detector-shared-problems","operation":"move","beforeSHA256":"c3fba1f619d9fa82c144218a2e5408f99396070490631920a72f7ead7b84d288","afterSHA256":"c3fba1f619d9fa82c144218a2e5408f99396070490631920a72f7ead7b84d288","reason":"Same source body in authored semantic page"},{"oldId":"b-0fe56de7f2ef4e88eaf3ff5b","targetId":"b-0fe56de7f2ef4e88eaf3ff5b","page":"dense-detector-shared-problems","operation":"move","beforeSHA256":"7d8fe2f5f510715c2b69ed15383f5797f594930db9819d67d883ae2d3c44722e","afterSHA256":"7d8fe2f5f510715c2b69ed15383f5797f594930db9819d67d883ae2d3c44722e","reason":"Same source body in authored semantic page"},{"oldId":"b-47af5d3609268774a299b4ff","targetId":"b-47af5d3609268774a299b4ff","page":"dense-detector-shared-problems","operation":"move","beforeSHA256":"159608d00faa18d3a91c3edd497cb1a181cb69965e0d2c41b1161b30140fa3ae","afterSHA256":"159608d00faa18d3a91c3edd497cb1a181cb69965e0d2c41b1161b30140fa3ae","reason":"Same source body in authored semantic page"},{"oldId":"b-ea0d64617d33b325ad2fd149","targetId":"b-ea0d64617d33b325ad2fd149","page":"dense-detector-shared-problems","operation":"move","beforeSHA256":"6b8c8150ee61580fb752c1fde1e79a70c46e4b0ffcd07319b820b1e4ce5c430e","afterSHA256":"6b8c8150ee61580fb752c1fde1e79a70c46e4b0ffcd07319b820b1e4ce5c430e","reason":"Same source body in authored semantic page"},{"oldId":"b-0dfd6cc71975d9fd709ae674","targetId":"b-0dfd6cc71975d9fd709ae674","page":"dense-detector-shared-problems","operation":"move","beforeSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","afterSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","reason":"Same source body in authored semantic page"},{"oldId":"b-73941df9cf56bf7300cec645","targetId":"b-73941df9cf56bf7300cec645","page":"dense-detector-shared-problems","operation":"move","beforeSHA256":"a3163e9b7618b50175088fce50e8c02bf38584dbcb466052e3f6745aa91110a4","afterSHA256":"a3163e9b7618b50175088fce50e8c02bf38584dbcb466052e3f6745aa91110a4","reason":"Same source body in authored semantic page"},{"oldId":"reference-proposal-detection-title","targetId":"b-2abd9790d16b5dea373a41cd","operation":"retain-source-identity"},{"oldId":"anchor-training-targets-title","targetId":"b-c4de52564e407b770f6ef417","operation":"retain-source-identity"},{"oldId":"anchor-training-targets-bridge","targetId":"b-181f205bd73dc87b9cbdbe35","operation":"retain-source-identity"},{"oldId":"detector-lab10-evidence-title","targetId":"b-aee752dd021bba3dd794d4ae","operation":"retain-source-identity"},{"oldId":"detector-lab10-evidence-bridge","targetId":"b-75f00b745057ddf5e18dd7e1","operation":"retain-source-identity"},{"oldId":"lab10-saved-failures","targetId":"b-982d9de3efb52a6c552b6b69","operation":"retain-source-identity"},{"oldId":"dense-detector-shared-problems-title","targetId":"b-6362773778e8ee0caabb0e39","operation":"retain-source-identity"},{"oldId":"dense-detector-shared-problems-bridge","targetId":"b-27a45f724955fef9bd94b76d","operation":"retain-source-identity"}],"blocks":[{"id":"b-2abd9790d16b5dea373a41cd","type":"h2","scope":"reference-proposal-detection","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-2abd9790d16b5dea373a41cd","content":"## 分开参考框、候选框与最终检测框","sha256":"17acdb3be6260bfa11ffa990cb1875a7a3114aa4d06e3c55a623f8af661c440f","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-186519f1b6f69e3db2ee6247","prerequisites":[{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-186519f1b6f69e3db2ee6247","type":"h2","scope":"explanation","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-186519f1b6f69e3db2ee6247","content":"## 解释、算例与证据","sha256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-2abd9790d16b5dea373a41cd","nextBlock":"b-20a30a24e745c5c7804b77bb","prerequisites":[{"id":"b-2abd9790d16b5dea373a41cd","url":"#b-2abd9790d16b5dea373a41cd","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-20a30a24e745c5c7804b77bb","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-20a30a24e745c5c7804b77bb","content":"# P32 · Faster R-CNN：让候选区域也成为预测","sha256":"fa5fee02eec78a0fc68677365db555ff86cf68a2947a56e2a2f54c2a789e96f2","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-186519f1b6f69e3db2ee6247","nextBlock":"b-4b845ae67f6bca64d7a9262d","prerequisites":[{"id":"b-186519f1b6f69e3db2ee6247","url":"#b-186519f1b6f69e3db2ee6247","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-4b845ae67f6bca64d7a9262d","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-4b845ae67f6bca64d7a9262d","content":"按“参考框→RPN 候选→最终框”给图中的矩形分配身份，并把人工真值单独留在监督侧。RPN 让候选变为可学习预测。LAB10 采用共享原理的一阶段教学变体，不应把它说成原版 Faster R-CNN。","sha256":"6b45cf47a23e9eec2f15663b9a1624db57adbc4f2a6825dd6bc21c96bfb93aa4","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-20a30a24e745c5c7804b77bb","nextBlock":"b-b41f25904357eb5693af08bf","prerequisites":[{"id":"b-20a30a24e745c5c7804b77bb","url":"#b-20a30a24e745c5c7804b77bb","role":"section-title"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-b41f25904357eb5693af08bf","type":"figure","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-b41f25904357eb5693af08bf","content":"图内标签与数值：Lecture 03 课堂图解 P32；共享 F；RPN；前景分数 + 框偏移；区域读出；分类 + 最终框修正；候选框 → 供后级使用；参考框；参考几何 → 预测候选 → 最终检测框\n\n共享 F 分为 RPN 和区域预测；RPN 产生候选并供后级使用，参考框与偏移后的候选用不同颜色表示。\n\n- 共享 F\n- RPN\n- 前景分数 + 框偏移\n- 区域读出\n- 分类 + 最终框修正\n- 候选框 → 供后级使用\n- 参考框\n- 参考几何 → 预测候选 → 最终检测框共享 F 分为 RPN 和区域预测；RPN 产生候选并供后级使用，参考框与偏移后的候选用不同颜色表示。 [查看对应图解状态](/#/l3-r2-p32/1)[查看完整图解](/course/figures/l3-r2-p32-1.html)","sha256":"682d9ad365a9a0ea524a22340299a5f3812904ef83a368c0161ca3325f532cec","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"共享 F 分为 RPN 和区域预测；RPN 产生候选并供后级使用，参考框与偏移后的候选用不同颜色表示。 查看对应图解状态","labels":["共享 F","RPN","前景分数 + 框偏移","区域读出","分类 + 最终框修正","候选框 → 供后级使用","参考框","参考几何 → 预测候选 → 最终检测框"],"description":"共享 F 分为 RPN 和区域预测；RPN 产生候选并供后级使用，参考框与偏移后的候选用不同颜色表示。共享 FRPN前景分数 + 框偏移区域读出分类 + 最终框修正候选框 → 供后级使用参考框参考几何 → 预测候选 → 最终检测框","evidenceIdentity":"共享 F 分为 RPN 和区域预测；RPN 产生候选并供后级使用，参考框与偏移后的候选用不同颜色表示。 查看对应图解状态","images":[{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""}]},"previousBlock":"b-4b845ae67f6bca64d7a9262d","nextBlock":"b-12220d5b859b781cdec0e1b7","prerequisites":[{"id":"b-20a30a24e745c5c7804b77bb","url":"#b-20a30a24e745c5c7804b77bb","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-12220d5b859b781cdec0e1b7","type":"h2","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-12220d5b859b781cdec0e1b7","content":"## 课堂讲述","sha256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b41f25904357eb5693af08bf","nextBlock":"b-dabd833266f6e275ec84f502","prerequisites":[{"id":"b-20a30a24e745c5c7804b77bb","url":"#b-20a30a24e745c5c7804b77bb","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-dabd833266f6e275ec84f502","type":"p","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-dabd833266f6e275ec84f502","content":"Faster R-CNN 增加了区域建议网络，也就是 RPN。它在共享特征上判断一些参考框是否像前景，并预测位置修正，形成后面使用的候选。\n请区分图上的三个对象：参考框是我们提供的几何起点，候选框是这一步的预测，最终检测框是后续处理后的结果。训练时，还要根据真实标注定义哪些参考框负责哪些对象。\n至此，从外部产生区域、反复编码区域，到共享卷积、再学习候选，瓶颈被一段段重新组织。这条路线有跨团队与合作者的汇合，不能把原始 R-CNN、SPP-net 和 Faster R-CNN 都归给一个人。它们相连的是具体问题与可复用的计算。","sha256":"febdeb92cc5c7d107b2125830e7f25e5103c9845e4447dd8ccda5f273e85c59b","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-12220d5b859b781cdec0e1b7","nextBlock":"b-bf6003d8695683a668235d24","prerequisites":[{"id":"b-12220d5b859b781cdec0e1b7","url":"#b-12220d5b859b781cdec0e1b7","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-bf6003d8695683a668235d24","type":"h3","scope":"rpn-three-boxes","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-bf6003d8695683a668235d24","content":"### 参考框、候选框与最终检测框","sha256":"8d8642c37d0a16aa2c09a67fd19a693edc3211397fb9026649867e38a51c165b","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":"rpn-three-boxes","formulas":[],"previousBlock":"b-dabd833266f6e275ec84f502","nextBlock":"b-bfb211a278f543d28f330665","prerequisites":[{"id":"b-12220d5b859b781cdec0e1b7","url":"#b-12220d5b859b781cdec0e1b7","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"},{"id":"b-bfb211a278f543d28f330665","url":"#b-bfb211a278f543d28f330665","relation":"same-reasoning"},{"id":"b-9700ffd63e3598027017fc2c","url":"#b-9700ffd63e3598027017fc2c","relation":"same-reasoning"},{"id":"b-3d61f626d5daf110bda0c610","url":"#b-3d61f626d5daf110bda0c610","relation":"same-reasoning"},{"id":"b-d05cd9d1f66d10cb8526492f","url":"#b-d05cd9d1f66d10cb8526492f","relation":"same-reasoning"}]},{"id":"b-bfb211a278f543d28f330665","type":"p","scope":"rpn-three-boxes","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-bfb211a278f543d28f330665","content":"RPN 在共享特征图上铺一组参考框。参考框由网格位置、预设尺度和长宽比决定；它们在看图之前就可生成，不能当成检测结果。网络为每个参考框输出物体性分数和四个偏移，经解码、筛选后形成 proposals，再交给区域头作更具体的类别判断和位置修正。","sha256":"ff6018e332450f304b746defc546d270e531208c2bd0354e7e32a269661363c1","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":"rpn-three-boxes","formulas":[],"previousBlock":"b-bf6003d8695683a668235d24","nextBlock":"b-9700ffd63e3598027017fc2c","prerequisites":[{"id":"b-bf6003d8695683a668235d24","url":"#b-bf6003d8695683a668235d24","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"},{"id":"b-bf6003d8695683a668235d24","url":"#b-bf6003d8695683a668235d24","relation":"same-reasoning"},{"id":"b-9700ffd63e3598027017fc2c","url":"#b-9700ffd63e3598027017fc2c","relation":"same-reasoning"},{"id":"b-3d61f626d5daf110bda0c610","url":"#b-3d61f626d5daf110bda0c610","relation":"same-reasoning"},{"id":"b-d05cd9d1f66d10cb8526492f","url":"#b-d05cd9d1f66d10cb8526492f","relation":"same-reasoning"}]},{"id":"b-9700ffd63e3598027017fc2c","type":"p","scope":"rpn-three-boxes","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-9700ffd63e3598027017fc2c","content":"因此画面中的三个矩形有不同身份：anchor 是几何起点，proposal 是 RPN 的预测，最终 box 是整个检测流程交付的结果。训练真值是第四种对象，它负责给预测规定目标，并不在推理时送进模型告诉它答案。","sha256":"ec5bbb4f56f3be17858a079f938ad06774395a07f9f46cda16ebf9c2538d0144","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":"rpn-three-boxes","formulas":[],"previousBlock":"b-bfb211a278f543d28f330665","nextBlock":"b-3d61f626d5daf110bda0c610","prerequisites":[{"id":"b-bf6003d8695683a668235d24","url":"#b-bf6003d8695683a668235d24","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"},{"id":"b-bf6003d8695683a668235d24","url":"#b-bf6003d8695683a668235d24","relation":"same-reasoning"},{"id":"b-bfb211a278f543d28f330665","url":"#b-bfb211a278f543d28f330665","relation":"same-reasoning"},{"id":"b-3d61f626d5daf110bda0c610","url":"#b-3d61f626d5daf110bda0c610","relation":"same-reasoning"},{"id":"b-d05cd9d1f66d10cb8526492f","url":"#b-d05cd9d1f66d10cb8526492f","relation":"same-reasoning"}]},{"id":"b-3d61f626d5daf110bda0c610","type":"p","scope":"rpn-three-boxes","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-3d61f626d5daf110bda0c610","content":"同一张 F 可同时供 RPN 与后级区域头使用。它把“在哪里值得再看”从外部算法变成从图像特征学习的任务。这是 Faster R-CNN 与 Fast R-CNN 相比发生改变的具体一段，不能只记一个更快的名字。","sha256":"87e8f6e090d6b2bb0be5d6005c6ef7fae57d2b05e1c4df5c825928472ed755bf","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":"rpn-three-boxes","formulas":[],"previousBlock":"b-9700ffd63e3598027017fc2c","nextBlock":"b-d05cd9d1f66d10cb8526492f","prerequisites":[{"id":"b-bf6003d8695683a668235d24","url":"#b-bf6003d8695683a668235d24","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"},{"id":"b-bf6003d8695683a668235d24","url":"#b-bf6003d8695683a668235d24","relation":"same-reasoning"},{"id":"b-bfb211a278f543d28f330665","url":"#b-bfb211a278f543d28f330665","relation":"same-reasoning"},{"id":"b-9700ffd63e3598027017fc2c","url":"#b-9700ffd63e3598027017fc2c","relation":"same-reasoning"},{"id":"b-d05cd9d1f66d10cb8526492f","url":"#b-d05cd9d1f66d10cb8526492f","relation":"same-reasoning"}]},{"id":"b-d05cd9d1f66d10cb8526492f","type":"p","scope":"rpn-three-boxes","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-d05cd9d1f66d10cb8526492f","content":"LAB10 采用单层锚框直接分类与回归的一阶段教学检测器，不是原版两阶段 RPN。它与本页共享锚框、匹配、位置损失等原理，保留了可运行的学习闭环；读实验时应辨明哪些流程共用、哪些结构不同。","sha256":"d39d8ade1a97f74de22f5548d14b4a3ca83a4cff593b2c987f0b8ead3d6a7d6e","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":"rpn-three-boxes","formulas":[],"previousBlock":"b-3d61f626d5daf110bda0c610","nextBlock":"b-eb65dd8842f5c37b38c6d9c9","prerequisites":[{"id":"b-bf6003d8695683a668235d24","url":"#b-bf6003d8695683a668235d24","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"},{"id":"b-bf6003d8695683a668235d24","url":"#b-bf6003d8695683a668235d24","relation":"same-reasoning"},{"id":"b-bfb211a278f543d28f330665","url":"#b-bfb211a278f543d28f330665","relation":"same-reasoning"},{"id":"b-9700ffd63e3598027017fc2c","url":"#b-9700ffd63e3598027017fc2c","relation":"same-reasoning"},{"id":"b-3d61f626d5daf110bda0c610","url":"#b-3d61f626d5daf110bda0c610","relation":"same-reasoning"}]},{"id":"b-eb65dd8842f5c37b38c6d9c9","type":"h4","scope":"reference-proposal-detection","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-eb65dd8842f5c37b38c6d9c9","content":"#### 图示条件与公式约定","sha256":"ef7f915c1ae63ed67bd23fecf050249273ba0efa8965d2c8d1c1c6b2d3178122","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d05cd9d1f66d10cb8526492f","nextBlock":"b-171e4ea2e1cf37b2ee0113ea","prerequisites":[{"id":"b-bf6003d8695683a668235d24","url":"#b-bf6003d8695683a668235d24","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-171e4ea2e1cf37b2ee0113ea","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-171e4ea2e1cf37b2ee0113ea","content":"参考框是几何起点，RPN 输出 objectness 与偏移形成 proposals，后级再预测类别与框。图省略 NMS、采样与正负匹配阈值，它们在补充材料中保留。","sha256":"bbfc204ffba300b74c47d93813472b503913a6c38c28d2a201fabd8502b3bb98","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-eb65dd8842f5c37b38c6d9c9","nextBlock":"b-b78d15849ee3edade6e23dca","prerequisites":[{"id":"b-eb65dd8842f5c37b38c6d9c9","url":"#b-eb65dd8842f5c37b38c6d9c9","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-b78d15849ee3edade6e23dca","type":"h4","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-b78d15849ee3edade6e23dca","content":"#### 这条线索的展开与后续阅读","sha256":"48b8229f7559b36ed4d5628416e3991e5c73293fbce43374db67386c8736f3bc","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-171e4ea2e1cf37b2ee0113ea","nextBlock":"b-3a30bc257b48623626999482","prerequisites":[{"id":"b-eb65dd8842f5c37b38c6d9c9","url":"#b-eb65dd8842f5c37b38c6d9c9","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-3a30bc257b48623626999482","type":"ul","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-3a30bc257b48623626999482","content":"- [候选区域，也从共享表示中学习](/supplement.html?return=l3-r2-p32/0#/l3-rpn/0) 本页说明\n- [RPN 如何把参考框变成训练样本｜问题从哪里出现](/supplement.html?return=l3-r2-p32/0#/l3-rpn-anchors-problem/0) 本页说明\n- [RPN 如何把参考框变成训练样本｜核心改动怎样工作](/supplement.html?return=l3-r2-p32/0#/l3-rpn-anchors-mechanism/0) 本页说明\n- [RPN 如何把参考框变成训练样本｜一起算一个完整例子](/supplement.html?return=l3-r2-p32/0#/l3-rpn-anchors-example/0) 本页说明\n- [RPN 如何把参考框变成训练样本｜怎样检查这项解释](/supplement.html?return=l3-r2-p32/0#/l3-rpn-anchors-evidence/0) 本页说明\n- [RPN 如何把参考框变成训练样本｜什么结论还不能推出](/supplement.html?return=l3-r2-p32/0#/l3-rpn-anchors-boundary/0) 本页说明\n- [一阶段检测把预测铺满特征图｜问题从哪里出现](/supplement.html?return=l3-r2-p32/0#/l3-dense-one-stage-problem/0) L03补充阅读\n- [一阶段检测把预测铺满特征图｜核心改动怎样工作](/supplement.html?return=l3-r2-p32/0#/l3-dense-one-stage-mechanism/0) L03补充阅读","sha256":"9907f492b45d33d2cd095cf93e4f2992b615967bc013d051ce64b24f87065754","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b78d15849ee3edade6e23dca","nextBlock":"b-0158a60b62fbc2e82f98bb50","prerequisites":[{"id":"b-b78d15849ee3edade6e23dca","url":"#b-b78d15849ee3edade6e23dca","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-0158a60b62fbc2e82f98bb50","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-0158a60b62fbc2e82f98bb50","content":"- [候选区域，也从共享表示中学习](/supplement.html?return=l3-r2-p32/0#/l3-rpn/0) 本页说明","sha256":"b509fe9b9654e59b2393a987864d15730dba547c20e00743ea434b4dd911fbc9","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3a30bc257b48623626999482","nextBlock":"b-2bfac8a38ce5d482cef7c0e9","prerequisites":[{"id":"b-b78d15849ee3edade6e23dca","url":"#b-b78d15849ee3edade6e23dca","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-2bfac8a38ce5d482cef7c0e9","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-2bfac8a38ce5d482cef7c0e9","content":"- [RPN 如何把参考框变成训练样本｜问题从哪里出现](/supplement.html?return=l3-r2-p32/0#/l3-rpn-anchors-problem/0) 本页说明","sha256":"19ab452623b51a31d2b44d4c7e9d38c8d7fd345a5d0d975abe5a42661bea965d","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-0158a60b62fbc2e82f98bb50","nextBlock":"b-16a7a5fda344ac718152903b","prerequisites":[{"id":"b-b78d15849ee3edade6e23dca","url":"#b-b78d15849ee3edade6e23dca","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-16a7a5fda344ac718152903b","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-16a7a5fda344ac718152903b","content":"- [RPN 如何把参考框变成训练样本｜核心改动怎样工作](/supplement.html?return=l3-r2-p32/0#/l3-rpn-anchors-mechanism/0) 本页说明","sha256":"84fada740d51b378cea3e0e47e6651c3bb1828c75c12514430355bdb75da67d4","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-2bfac8a38ce5d482cef7c0e9","nextBlock":"b-2c8e8ec10d779a7959ac5235","prerequisites":[{"id":"b-b78d15849ee3edade6e23dca","url":"#b-b78d15849ee3edade6e23dca","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-2c8e8ec10d779a7959ac5235","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-2c8e8ec10d779a7959ac5235","content":"- [RPN 如何把参考框变成训练样本｜一起算一个完整例子](/supplement.html?return=l3-r2-p32/0#/l3-rpn-anchors-example/0) 本页说明","sha256":"b7fb3753c687fe3752b886d2f752f2ab90796048541029814935ab403f4031e4","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-16a7a5fda344ac718152903b","nextBlock":"b-71ecab44188d91b44d011801","prerequisites":[{"id":"b-b78d15849ee3edade6e23dca","url":"#b-b78d15849ee3edade6e23dca","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-71ecab44188d91b44d011801","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-71ecab44188d91b44d011801","content":"- [RPN 如何把参考框变成训练样本｜怎样检查这项解释](/supplement.html?return=l3-r2-p32/0#/l3-rpn-anchors-evidence/0) 本页说明","sha256":"91124d0e21521d83268cc3ffa1876dac09f5748aa130e072ee53b1a27cb9e450","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-2c8e8ec10d779a7959ac5235","nextBlock":"b-d1dd5dac3ea6e2fb9d4f236e","prerequisites":[{"id":"b-b78d15849ee3edade6e23dca","url":"#b-b78d15849ee3edade6e23dca","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-d1dd5dac3ea6e2fb9d4f236e","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-d1dd5dac3ea6e2fb9d4f236e","content":"- [RPN 如何把参考框变成训练样本｜什么结论还不能推出](/supplement.html?return=l3-r2-p32/0#/l3-rpn-anchors-boundary/0) 本页说明","sha256":"5a6614c9865a2a725016da4bdb0f7756f578db6952ab280a0accbeb12cb5c7c8","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-71ecab44188d91b44d011801","nextBlock":"b-c5b9177a9a22aeb431cb2791","prerequisites":[{"id":"b-b78d15849ee3edade6e23dca","url":"#b-b78d15849ee3edade6e23dca","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-c5b9177a9a22aeb431cb2791","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-c5b9177a9a22aeb431cb2791","content":"- [一阶段检测把预测铺满特征图｜问题从哪里出现](/supplement.html?return=l3-r2-p32/0#/l3-dense-one-stage-problem/0) L03补充阅读","sha256":"4cf00996a0f6282ac8d4f2c1ef4f559764812bb995da338db000ec5a7e6c1330","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d1dd5dac3ea6e2fb9d4f236e","nextBlock":"b-acfe7f50e1e8b461a2a75a59","prerequisites":[{"id":"b-b78d15849ee3edade6e23dca","url":"#b-b78d15849ee3edade6e23dca","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-acfe7f50e1e8b461a2a75a59","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-acfe7f50e1e8b461a2a75a59","content":"- [一阶段检测把预测铺满特征图｜核心改动怎样工作](/supplement.html?return=l3-r2-p32/0#/l3-dense-one-stage-mechanism/0) L03补充阅读","sha256":"b464b01fa76793dd871c621138c876cd92257e297f945300d9dbbd5206e92eac","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reference-proposal-detection","title":"分开参考框、候选框与最终检测框","number":1},"sectionTitle":"分开参考框、候选框与最终检测框","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c5b9177a9a22aeb431cb2791","nextBlock":null,"prerequisites":[{"id":"b-b78d15849ee3edade6e23dca","url":"#b-b78d15849ee3edade6e23dca","role":"section-title"},{"id":"b-4b845ae67f6bca64d7a9262d","url":"#b-4b845ae67f6bca64d7a9262d","role":"problem-context"}],"relatedBlocks":[{"id":"b-b41f25904357eb5693af08bf","url":"#b-b41f25904357eb5693af08bf","relation":"figure-and-explanation"}]},{"id":"b-c4de52564e407b770f6ef417","type":"h2","scope":"anchor-training-targets","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-c4de52564e407b770f6ef417","content":"## 给正例、背景和忽略项分配任务","sha256":"cfdc62b942e00569aee310fba942c1038f74c2bd186827274b1fd3d010a95bd6","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"anchor-training-targets","title":"给正例、背景和忽略项分配任务","number":2},"sectionTitle":"给正例、背景和忽略项分配任务","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32","l3-r2-p33"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-181f205bd73dc87b9cbdbe35","prerequisites":[{"id":"b-181f205bd73dc87b9cbdbe35","url":"#b-181f205bd73dc87b9cbdbe35","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-181f205bd73dc87b9cbdbe35","type":"p","scope":"anchor-training-targets","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-181f205bd73dc87b9cbdbe35","content":"先明确训练分配，才能知道哪一项预测拥有位置目标。","sha256":"2660a6ba005bfad3a1f1ff72dea53a40cec39672de6d4dde06f5c942691ccf5d","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"anchor-training-targets","title":"给正例、背景和忽略项分配任务","number":2},"sectionTitle":"给正例、背景和忽略项分配任务","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32","l3-r2-p33"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c4de52564e407b770f6ef417","nextBlock":"b-9bad1749ccd77ae4fb1e627c","prerequisites":[{"id":"b-c4de52564e407b770f6ef417","url":"#b-c4de52564e407b770f6ef417","role":"section-title"}],"relatedBlocks":[]},{"id":"b-9bad1749ccd77ae4fb1e627c","type":"h3","scope":"anchor-training-assignment","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-9bad1749ccd77ae4fb1e627c","content":"### 为什么要有正例、背景和忽略项","sha256":"7af001429caa370b259fa0b7820036e8a5a1052d83c445a1e2ab6fbbda8c7c93","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"anchor-training-targets","title":"给正例、背景和忽略项分配任务","number":2},"sectionTitle":"为什么要有正例、背景和忽略项","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32","l3-r2-p33"],"knowledgeBlock":"anchor-training-assignment","formulas":[],"previousBlock":"b-181f205bd73dc87b9cbdbe35","nextBlock":"b-94169f26c39be5ca7ebae497","prerequisites":[{"id":"b-c4de52564e407b770f6ef417","url":"#b-c4de52564e407b770f6ef417","role":"section-title"},{"id":"b-181f205bd73dc87b9cbdbe35","url":"#b-181f205bd73dc87b9cbdbe35","role":"problem-context"}],"relatedBlocks":[{"id":"b-94169f26c39be5ca7ebae497","url":"#b-94169f26c39be5ca7ebae497","relation":"same-reasoning"},{"id":"b-d80f2878b651375aec18736e","url":"#b-d80f2878b651375aec18736e","relation":"same-reasoning"},{"id":"b-294c9d12e24bab16c72b8945","url":"#b-294c9d12e24bab16c72b8945","relation":"same-reasoning"},{"id":"b-522412585d246c7a10026eb2","url":"#b-522412585d246c7a10026eb2","relation":"same-reasoning"}]},{"id":"b-94169f26c39be5ca7ebae497","type":"p","scope":"anchor-training-assignment","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-94169f26c39be5ca7ebae497","content":"给定参考框与真值的 IoU 矩阵，匹配规则把参考框变成训练样本。例如课堂教学阈值若设正例至少 0.7、背景小于 0.3，则 IoU 为 0.8、0.4、0.1 的三个框分别是正例、忽略、背景。只有正例有对应的框回归目标。忽略的用途是暂不把模糊覆盖强行当正或负，而不是悄悄改成背景。","sha256":"290dba49c527e68e2aba4a2ee0b4d28e626d8e489edab4605ea5a235e8ea7fb2","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"anchor-training-targets","title":"给正例、背景和忽略项分配任务","number":2},"sectionTitle":"为什么要有正例、背景和忽略项","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32","l3-r2-p33"],"knowledgeBlock":"anchor-training-assignment","formulas":[],"previousBlock":"b-9bad1749ccd77ae4fb1e627c","nextBlock":"b-d80f2878b651375aec18736e","prerequisites":[{"id":"b-9bad1749ccd77ae4fb1e627c","url":"#b-9bad1749ccd77ae4fb1e627c","role":"section-title"},{"id":"b-181f205bd73dc87b9cbdbe35","url":"#b-181f205bd73dc87b9cbdbe35","role":"problem-context"}],"relatedBlocks":[{"id":"b-9bad1749ccd77ae4fb1e627c","url":"#b-9bad1749ccd77ae4fb1e627c","relation":"same-reasoning"},{"id":"b-d80f2878b651375aec18736e","url":"#b-d80f2878b651375aec18736e","relation":"same-reasoning"},{"id":"b-294c9d12e24bab16c72b8945","url":"#b-294c9d12e24bab16c72b8945","relation":"same-reasoning"},{"id":"b-522412585d246c7a10026eb2","url":"#b-522412585d246c7a10026eb2","relation":"same-reasoning"}]},{"id":"b-d80f2878b651375aec18736e","type":"p","scope":"anchor-training-assignment","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-d80f2878b651375aec18736e","content":"真实 LAB10 使用另一组公开写明的阈值：正例至少 0.45，背景低于 0.25，中间忽略；再用逐次最大 IoU 为每个真值保留一个不同 anchor，避免某个对象完全收不到位置监督。强制匹配服务于学习，评价时不能为了“每个对象都匹配一次”照搬这条规则。","sha256":"9950be4406a476d8d766763df53ea513bbe7eed02f34aff4f837c30a3d045b5c","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"anchor-training-targets","title":"给正例、背景和忽略项分配任务","number":2},"sectionTitle":"为什么要有正例、背景和忽略项","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32","l3-r2-p33"],"knowledgeBlock":"anchor-training-assignment","formulas":[],"previousBlock":"b-94169f26c39be5ca7ebae497","nextBlock":"b-294c9d12e24bab16c72b8945","prerequisites":[{"id":"b-9bad1749ccd77ae4fb1e627c","url":"#b-9bad1749ccd77ae4fb1e627c","role":"section-title"},{"id":"b-181f205bd73dc87b9cbdbe35","url":"#b-181f205bd73dc87b9cbdbe35","role":"problem-context"}],"relatedBlocks":[{"id":"b-9bad1749ccd77ae4fb1e627c","url":"#b-9bad1749ccd77ae4fb1e627c","relation":"same-reasoning"},{"id":"b-94169f26c39be5ca7ebae497","url":"#b-94169f26c39be5ca7ebae497","relation":"same-reasoning"},{"id":"b-294c9d12e24bab16c72b8945","url":"#b-294c9d12e24bab16c72b8945","relation":"same-reasoning"},{"id":"b-522412585d246c7a10026eb2","url":"#b-522412585d246c7a10026eb2","relation":"same-reasoning"}]},{"id":"b-294c9d12e24bab16c72b8945","type":"p","scope":"anchor-training-assignment","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-294c9d12e24bab16c72b8945","content":"空图同样是有效训练输入。所有参考框都应学背景，回归项没有正例便归零。如果只训练总有物体的图片，模型可能从未被要求学会空输出。","sha256":"157cb5eec4f8576589386132685fedffe3cc7d7ba664168060367d6e5a58232b","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"anchor-training-targets","title":"给正例、背景和忽略项分配任务","number":2},"sectionTitle":"为什么要有正例、背景和忽略项","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32","l3-r2-p33"],"knowledgeBlock":"anchor-training-assignment","formulas":[],"previousBlock":"b-d80f2878b651375aec18736e","nextBlock":"b-522412585d246c7a10026eb2","prerequisites":[{"id":"b-9bad1749ccd77ae4fb1e627c","url":"#b-9bad1749ccd77ae4fb1e627c","role":"section-title"},{"id":"b-181f205bd73dc87b9cbdbe35","url":"#b-181f205bd73dc87b9cbdbe35","role":"problem-context"}],"relatedBlocks":[{"id":"b-9bad1749ccd77ae4fb1e627c","url":"#b-9bad1749ccd77ae4fb1e627c","relation":"same-reasoning"},{"id":"b-94169f26c39be5ca7ebae497","url":"#b-94169f26c39be5ca7ebae497","relation":"same-reasoning"},{"id":"b-d80f2878b651375aec18736e","url":"#b-d80f2878b651375aec18736e","relation":"same-reasoning"},{"id":"b-522412585d246c7a10026eb2","url":"#b-522412585d246c7a10026eb2","relation":"same-reasoning"}]},{"id":"b-522412585d246c7a10026eb2","type":"p","scope":"anchor-training-assignment","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-522412585d246c7a10026eb2","content":"负例远多于正例时，本实验选难负例参与分类损失，最多正例数的三倍、每图至少十二个。阅读代码时应看 loss 的掩码与归一化，而不只看 $L=L_{cls}+\\lambda L_{box}$ 这一行总公式。","sha256":"767e9cdba079c3d5360d01ca6637ff47b625f805ea988b0a945b3600317d283c","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"anchor-training-targets","title":"给正例、背景和忽略项分配任务","number":2},"sectionTitle":"为什么要有正例、背景和忽略项","currentQuestion":"候选、训练匹配和框偏移分别承担哪一步？","sourceSlides":["l3-r2-p32","l3-r2-p33"],"knowledgeBlock":"anchor-training-assignment","formulas":[{"latex":"L=L_{cls}+\\lambda L_{box}","definitionContext":"给定参考框与真值的 IoU 矩阵，匹配规则把参考框变成训练样本。例如课堂教学阈值若设正例至少 0.7、背景小于 0.3，则 IoU 为 0.8、0.4、0.1 的三个框分别是正例、忽略、背景。只有正例有对应的框回归目标。忽略的用途是暂不把模糊覆盖强行当正或负，而不是悄悄改成背景。\n真实 LAB10 使用另一组公开写明的阈值：正例至少 0.45，背景低于 0.25，中间忽略；再用逐次最大 IoU 为每个真值保留一个不同 anchor，避免某个对象完全收不到位置监督。强制匹配服务于学习，评价时不能为了“每个对象都匹配一次”照搬这条规则。\n空图同样是有效训练输入。所有参考框都应学背景，回归项没有正例便归零。如果只训练总有物体的图片，模型可能从未被要求学会空输出。\n负例远多于正例时，本实验选难负例参与分类损失，最多正例数的三倍、每图至少十二个。阅读代码时应看 loss 的掩码与归一化，而不只看 $L=L_{cls}+\\lambda L_{box}$ 这一行总公式。"}],"previousBlock":"b-294c9d12e24bab16c72b8945","nextBlock":null,"prerequisites":[{"id":"b-9bad1749ccd77ae4fb1e627c","url":"#b-9bad1749ccd77ae4fb1e627c","role":"section-title"},{"id":"b-181f205bd73dc87b9cbdbe35","url":"#b-181f205bd73dc87b9cbdbe35","role":"problem-context"}],"relatedBlocks":[{"id":"b-9bad1749ccd77ae4fb1e627c","url":"#b-9bad1749ccd77ae4fb1e627c","relation":"same-reasoning"},{"id":"b-94169f26c39be5ca7ebae497","url":"#b-94169f26c39be5ca7ebae497","relation":"same-reasoning"},{"id":"b-d80f2878b651375aec18736e","url":"#b-d80f2878b651375aec18736e","relation":"same-reasoning"},{"id":"b-294c9d12e24bab16c72b8945","url":"#b-294c9d12e24bab16c72b8945","relation":"same-reasoning"}]},{"id":"b-aee752dd021bba3dd794d4ae","type":"h2","scope":"detector-lab10-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-aee752dd021bba3dd794d4ae","content":"## 进入 LAB10：先读失败图，再解释高分","sha256":"860b9a40340598e6d1a5727ac35d9d740d417312dc1a93f0b7bd8c27ca1e6555","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"detector-lab10-evidence","title":"进入 LAB10：先读失败图，再解释高分","number":3},"sectionTitle":"进入 LAB10：先读失败图，再解释高分","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-75f00b745057ddf5e18dd7e1","prerequisites":[{"id":"b-75f00b745057ddf5e18dd7e1","url":"#b-75f00b745057ddf5e18dd7e1","role":"problem-context"}],"relatedBlocks":[{"id":"b-982d9de3efb52a6c552b6b69","url":"#b-982d9de3efb52a6c552b6b69","relation":"figure-and-explanation"}]},{"id":"b-75f00b745057ddf5e18dd7e1","type":"p","scope":"detector-lab10-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-75f00b745057ddf5e18dd7e1","content":"把错误分类、IoU 和 AP 带回已保存的图像到框学习记录。","sha256":"26f7a0a93a426c004ddf59454d530de63629ca6fdbd6b1c43e3de27003f50a69","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"detector-lab10-evidence","title":"进入 LAB10：先读失败图，再解释高分","number":3},"sectionTitle":"进入 LAB10：先读失败图，再解释高分","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-aee752dd021bba3dd794d4ae","nextBlock":"b-982d9de3efb52a6c552b6b69","prerequisites":[{"id":"b-aee752dd021bba3dd794d4ae","url":"#b-aee752dd021bba3dd794d4ae","role":"section-title"}],"relatedBlocks":[{"id":"b-982d9de3efb52a6c552b6b69","url":"#b-982d9de3efb52a6c552b6b69","relation":"figure-and-explanation"}]},{"id":"b-982d9de3efb52a6c552b6b69","type":"figure","scope":"detector-lab10-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-982d9de3efb52a6c552b6b69","content":"![LAB10 原失败图的样本35完整面板，绿色为真值，黄色为分数至少0.4的预测](/course/semantic-assets/lab10-sample35.png)已有实验记录的样本35面板：绿色框为真值，黄色框为分数至少0.4的预测。样本35的真值与预测各有两个框，但边界没有正确对应两个相邻对象。样本按几何失败选取，不能代表全部样本的错误比例。[查看完整失败图](/assets/display/88df211a98edfb7b00515441.webp)","sha256":"71eb0f8c9609aea79a42f8bde04d9ef244bbb12b1235d215122aec883d963246","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"detector-lab10-evidence","title":"进入 LAB10：先读失败图，再解释高分","number":3},"sectionTitle":"进入 LAB10：先读失败图，再解释高分","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"已有实验记录的样本35面板：绿色框为真值，黄色框为分数至少0.4的预测。样本35的真值与预测各有两个框，但边界没有正确对应两个相邻对象。样本按几何失败选取，不能代表全部样本的错误比例。查看完整失败图","labels":[],"description":"LAB10 原失败图的样本35完整面板，绿色为真值，黄色为分数至少0.4的预测","evidenceIdentity":"已有实验记录的样本35面板：绿色框为真值，黄色框为分数至少0.4的预测。样本35的真值与预测各有两个框，但边界没有正确对应两个相邻对象。样本按几何失败选取，不能代表全部样本的错误比例。查看完整失败图","images":[{"src":"/course/semantic-assets/lab10-sample35.png","alt":"LAB10 原失败图的样本35完整面板，绿色为真值，黄色为分数至少0.4的预测"}]},"previousBlock":"b-75f00b745057ddf5e18dd7e1","nextBlock":"b-ecc4865a85ab9a1ade3eb319","prerequisites":[{"id":"b-aee752dd021bba3dd794d4ae","url":"#b-aee752dd021bba3dd794d4ae","role":"section-title"},{"id":"b-75f00b745057ddf5e18dd7e1","url":"#b-75f00b745057ddf5e18dd7e1","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-ecc4865a85ab9a1ade3eb319","type":"h3","scope":"detector-learning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-ecc4865a85ab9a1ade3eb319","content":"### 从 RGB 到框：当前 LAB10 已经具备完整闭环","sha256":"76f6a9ca0e69cd8094fa23bd7ccbf42fb6bb705fb6fa0719bfe6760e58a0769f","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"detector-lab10-evidence","title":"进入 LAB10：先读失败图，再解释高分","number":3},"sectionTitle":"从 RGB 到框：当前 LAB10 已经具备完整闭环","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":"detector-learning-evidence","formulas":[],"previousBlock":"b-982d9de3efb52a6c552b6b69","nextBlock":"b-cb691a0fc8abcef8054b126e","prerequisites":[{"id":"b-aee752dd021bba3dd794d4ae","url":"#b-aee752dd021bba3dd794d4ae","role":"section-title"},{"id":"b-75f00b745057ddf5e18dd7e1","url":"#b-75f00b745057ddf5e18dd7e1","role":"problem-context"}],"relatedBlocks":[{"id":"b-982d9de3efb52a6c552b6b69","url":"#b-982d9de3efb52a6c552b6b69","relation":"figure-and-explanation"},{"id":"b-cb691a0fc8abcef8054b126e","url":"#b-cb691a0fc8abcef8054b126e","relation":"same-reasoning"},{"id":"b-447067e18fcf166999ea4137","url":"#b-447067e18fcf166999ea4137","relation":"same-reasoning"},{"id":"b-5b30369ff133e6b8dabd80a3","url":"#b-5b30369ff133e6b8dabd80a3","relation":"same-reasoning"},{"id":"b-dda7233a52b87fa2bbcd5d55","url":"#b-dda7233a52b87fa2bbcd5d55","relation":"same-reasoning"},{"id":"b-b922ccb95bec0009002c3d4b","url":"#b-b922ccb95bec0009002c3d4b","relation":"same-reasoning"}]},{"id":"b-cb691a0fc8abcef8054b126e","type":"p","scope":"detector-learning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-cb691a0fc8abcef8054b126e","content":"当前 LAB10 的主实验输入是真正的 48×48 RGB 合成图，目标是零至三个对象的类别与归一化框。卷积骨干生成 12×12 网格，每格两个参考框，共 288 个；两个头输出 `[B,288,3]` 类别 logits 与 `[B,288,4]` 偏移。图像进入 `model(images)`，坐标只参与监督与评价。","sha256":"2721b637cb6f94558f6713c359032adafb78e252bc13f52e76ce70874291e64b","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"detector-lab10-evidence","title":"进入 LAB10：先读失败图，再解释高分","number":3},"sectionTitle":"从 RGB 到框：当前 LAB10 已经具备完整闭环","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":"detector-learning-evidence","formulas":[],"previousBlock":"b-ecc4865a85ab9a1ade3eb319","nextBlock":"b-447067e18fcf166999ea4137","prerequisites":[{"id":"b-ecc4865a85ab9a1ade3eb319","url":"#b-ecc4865a85ab9a1ade3eb319","role":"section-title"},{"id":"b-75f00b745057ddf5e18dd7e1","url":"#b-75f00b745057ddf5e18dd7e1","role":"problem-context"}],"relatedBlocks":[{"id":"b-982d9de3efb52a6c552b6b69","url":"#b-982d9de3efb52a6c552b6b69","relation":"figure-and-explanation"},{"id":"b-ecc4865a85ab9a1ade3eb319","url":"#b-ecc4865a85ab9a1ade3eb319","relation":"same-reasoning"},{"id":"b-447067e18fcf166999ea4137","url":"#b-447067e18fcf166999ea4137","relation":"same-reasoning"},{"id":"b-5b30369ff133e6b8dabd80a3","url":"#b-5b30369ff133e6b8dabd80a3","relation":"same-reasoning"},{"id":"b-dda7233a52b87fa2bbcd5d55","url":"#b-dda7233a52b87fa2bbcd5d55","relation":"same-reasoning"},{"id":"b-b922ccb95bec0009002c3d4b","url":"#b-b922ccb95bec0009002c3d4b","relation":"same-reasoning"}]},{"id":"b-447067e18fcf166999ea4137","type":"p","scope":"detector-learning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-447067e18fcf166999ea4137","content":"保留的 2026-09-22 本地完整运行用 768 张训练图、96 张验证图、192 张保留图，训练 28 epoch。它以 TinySSD 的锚框检测闭环为起点，采用单特征层、两种参考尺度，不能称为 D2L 全部多尺度结构的复现。","sha256":"7f64f21b960df2c59e03e9c18028330d1e9a641549b00585d29c5ca6e0f9b5b3","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"detector-lab10-evidence","title":"进入 LAB10：先读失败图，再解释高分","number":3},"sectionTitle":"从 RGB 到框：当前 LAB10 已经具备完整闭环","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":"detector-learning-evidence","formulas":[],"previousBlock":"b-cb691a0fc8abcef8054b126e","nextBlock":"b-5b30369ff133e6b8dabd80a3","prerequisites":[{"id":"b-ecc4865a85ab9a1ade3eb319","url":"#b-ecc4865a85ab9a1ade3eb319","role":"section-title"},{"id":"b-75f00b745057ddf5e18dd7e1","url":"#b-75f00b745057ddf5e18dd7e1","role":"problem-context"}],"relatedBlocks":[{"id":"b-982d9de3efb52a6c552b6b69","url":"#b-982d9de3efb52a6c552b6b69","relation":"figure-and-explanation"},{"id":"b-ecc4865a85ab9a1ade3eb319","url":"#b-ecc4865a85ab9a1ade3eb319","relation":"same-reasoning"},{"id":"b-cb691a0fc8abcef8054b126e","url":"#b-cb691a0fc8abcef8054b126e","relation":"same-reasoning"},{"id":"b-5b30369ff133e6b8dabd80a3","url":"#b-5b30369ff133e6b8dabd80a3","relation":"same-reasoning"},{"id":"b-dda7233a52b87fa2bbcd5d55","url":"#b-dda7233a52b87fa2bbcd5d55","relation":"same-reasoning"},{"id":"b-b922ccb95bec0009002c3d4b","url":"#b-b922ccb95bec0009002c3d4b","relation":"same-reasoning"}]},{"id":"b-5b30369ff133e6b8dabd80a3","type":"table","scope":"detector-learning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-5b30369ff133e6b8dabd80a3","content":"| 条件 | AP50 | AP75 |\n| --- | --- | --- |\n| 随机初始化 | .00746 | .00003 |\n| 训练后原尺度 | .97303 | .70367 |\n| 同权重，仅对象尺寸减半 | .61490 | .17987 |","sha256":"ea3f2b25b776460f4e0e045b665251387ea0450a36a4ff6475b9f0188e257153","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"detector-lab10-evidence","title":"进入 LAB10：先读失败图，再解释高分","number":3},"sectionTitle":"从 RGB 到框：当前 LAB10 已经具备完整闭环","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":"detector-learning-evidence","formulas":[],"table":{"headers":["条件","AP50","AP75"],"rows":[["条件","AP50","AP75"],["随机初始化",".00746",".00003"],["训练后原尺度",".97303",".70367"],["同权重，仅对象尺寸减半",".61490",".17987"]]},"previousBlock":"b-447067e18fcf166999ea4137","nextBlock":"b-dda7233a52b87fa2bbcd5d55","prerequisites":[{"id":"b-ecc4865a85ab9a1ade3eb319","url":"#b-ecc4865a85ab9a1ade3eb319","role":"section-title"},{"id":"b-75f00b745057ddf5e18dd7e1","url":"#b-75f00b745057ddf5e18dd7e1","role":"problem-context"}],"relatedBlocks":[{"id":"b-982d9de3efb52a6c552b6b69","url":"#b-982d9de3efb52a6c552b6b69","relation":"figure-and-explanation"},{"id":"b-ecc4865a85ab9a1ade3eb319","url":"#b-ecc4865a85ab9a1ade3eb319","relation":"same-reasoning"},{"id":"b-cb691a0fc8abcef8054b126e","url":"#b-cb691a0fc8abcef8054b126e","relation":"same-reasoning"},{"id":"b-447067e18fcf166999ea4137","url":"#b-447067e18fcf166999ea4137","relation":"same-reasoning"},{"id":"b-dda7233a52b87fa2bbcd5d55","url":"#b-dda7233a52b87fa2bbcd5d55","relation":"same-reasoning"},{"id":"b-b922ccb95bec0009002c3d4b","url":"#b-b922ccb95bec0009002c3d4b","relation":"same-reasoning"}]},{"id":"b-dda7233a52b87fa2bbcd5d55","type":"p","scope":"detector-learning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-dda7233a52b87fa2bbcd5d55","content":"这些是已有运行记录，本轮内容整理没有重跑实验。原分布明显改善，且存在多对象与空图，支持“模型学习到了图像到对象位置的映射”。更严格位置门槛与小尺度条件仍显著下降，帮助定位下一步研究问题。红色与矩形、蓝色与椭圆相关，因此不能据此宣称自然图像泛化。","sha256":"8205b30be23a99f3c1540701846c79ab37442feeb48cf7193d02963131b7f9ab","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"detector-lab10-evidence","title":"进入 LAB10：先读失败图，再解释高分","number":3},"sectionTitle":"从 RGB 到框：当前 LAB10 已经具备完整闭环","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":"detector-learning-evidence","formulas":[],"previousBlock":"b-5b30369ff133e6b8dabd80a3","nextBlock":"b-b922ccb95bec0009002c3d4b","prerequisites":[{"id":"b-ecc4865a85ab9a1ade3eb319","url":"#b-ecc4865a85ab9a1ade3eb319","role":"section-title"},{"id":"b-75f00b745057ddf5e18dd7e1","url":"#b-75f00b745057ddf5e18dd7e1","role":"problem-context"}],"relatedBlocks":[{"id":"b-982d9de3efb52a6c552b6b69","url":"#b-982d9de3efb52a6c552b6b69","relation":"figure-and-explanation"},{"id":"b-ecc4865a85ab9a1ade3eb319","url":"#b-ecc4865a85ab9a1ade3eb319","relation":"same-reasoning"},{"id":"b-cb691a0fc8abcef8054b126e","url":"#b-cb691a0fc8abcef8054b126e","relation":"same-reasoning"},{"id":"b-447067e18fcf166999ea4137","url":"#b-447067e18fcf166999ea4137","relation":"same-reasoning"},{"id":"b-5b30369ff133e6b8dabd80a3","url":"#b-5b30369ff133e6b8dabd80a3","relation":"same-reasoning"},{"id":"b-b922ccb95bec0009002c3d4b","url":"#b-b922ccb95bec0009002c3d4b","relation":"same-reasoning"}]},{"id":"b-b922ccb95bec0009002c3d4b","type":"p","scope":"detector-learning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-b922ccb95bec0009002c3d4b","content":"读者可先看同一行的 RGB、真值框、预测框和失败图，再看匹配与损失，最后按需运行。原四维逻辑回归保留在热身位置，它说明分类更新，不承担检测证据。","sha256":"1caaa88a4dc7420a22d2566dc2b1a3e78a78af9d5f019befb84d7a50e0d3a296","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"detector-lab10-evidence","title":"进入 LAB10：先读失败图，再解释高分","number":3},"sectionTitle":"从 RGB 到框：当前 LAB10 已经具备完整闭环","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":"detector-learning-evidence","formulas":[],"previousBlock":"b-dda7233a52b87fa2bbcd5d55","nextBlock":null,"prerequisites":[{"id":"b-ecc4865a85ab9a1ade3eb319","url":"#b-ecc4865a85ab9a1ade3eb319","role":"section-title"},{"id":"b-75f00b745057ddf5e18dd7e1","url":"#b-75f00b745057ddf5e18dd7e1","role":"problem-context"}],"relatedBlocks":[{"id":"b-982d9de3efb52a6c552b6b69","url":"#b-982d9de3efb52a6c552b6b69","relation":"figure-and-explanation"},{"id":"b-ecc4865a85ab9a1ade3eb319","url":"#b-ecc4865a85ab9a1ade3eb319","relation":"same-reasoning"},{"id":"b-cb691a0fc8abcef8054b126e","url":"#b-cb691a0fc8abcef8054b126e","relation":"same-reasoning"},{"id":"b-447067e18fcf166999ea4137","url":"#b-447067e18fcf166999ea4137","relation":"same-reasoning"},{"id":"b-5b30369ff133e6b8dabd80a3","url":"#b-5b30369ff133e6b8dabd80a3","relation":"same-reasoning"},{"id":"b-dda7233a52b87fa2bbcd5d55","url":"#b-dda7233a52b87fa2bbcd5d55","relation":"same-reasoning"}]},{"id":"b-6362773778e8ee0caabb0e39","type":"h2","scope":"dense-detector-shared-problems","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-6362773778e8ee0caabb0e39","content":"## 一阶段检测简化哪一段，保留哪些问题？","sha256":"d61c9071fb532f6486e836e4aefcd33a67b270c9ab240829e82cbfb293c42a31","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-training-extensions","title":"选读：输出组织与训练分配还可以怎样改变？"},"page":{"id":"dense-detector-shared-problems","title":"一阶段检测简化哪一段，保留哪些问题？","number":4},"sectionTitle":"一阶段检测简化哪一段，保留哪些问题？","currentQuestion":"在已掌握区域、像素和实例后，按问题阅读结构与匹配扩展。","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-27a45f724955fef9bd94b76d","prerequisites":[{"id":"b-27a45f724955fef9bd94b76d","url":"#b-27a45f724955fef9bd94b76d","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-27a45f724955fef9bd94b76d","type":"p","scope":"dense-detector-shared-problems","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-27a45f724955fef9bd94b76d","content":"回查 LAB10 实现时，用这一页辨认它与两阶段路线共有和不同的部分。","sha256":"ab2c447549c789076478eea86511ff036f894ae246028db462d5d55ae4c0bf42","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-training-extensions","title":"选读：输出组织与训练分配还可以怎样改变？"},"page":{"id":"dense-detector-shared-problems","title":"一阶段检测简化哪一段，保留哪些问题？","number":4},"sectionTitle":"一阶段检测简化哪一段，保留哪些问题？","currentQuestion":"在已掌握区域、像素和实例后，按问题阅读结构与匹配扩展。","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6362773778e8ee0caabb0e39","nextBlock":"b-5aeffabcf350b1df8c2f3e65","prerequisites":[{"id":"b-6362773778e8ee0caabb0e39","url":"#b-6362773778e8ee0caabb0e39","role":"section-title"}],"relatedBlocks":[]},{"id":"b-5aeffabcf350b1df8c2f3e65","type":"h3","scope":"dense-detector-ssd-extension","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-5aeffabcf350b1df8c2f3e65","content":"### 选读：一阶段检测保留了哪些相同问题","sha256":"6d7a9f0134bacd810263ef2e018b0b9caefa60be8a73e1c4d75a595f843edc73","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-training-extensions","title":"选读：输出组织与训练分配还可以怎样改变？"},"page":{"id":"dense-detector-shared-problems","title":"一阶段检测简化哪一段，保留哪些问题？","number":4},"sectionTitle":"选读：一阶段检测保留了哪些相同问题","currentQuestion":"在已掌握区域、像素和实例后，按问题阅读结构与匹配扩展。","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":"dense-detector-ssd-extension","formulas":[],"previousBlock":"b-27a45f724955fef9bd94b76d","nextBlock":"b-5e05bacb7c023a91ee20a16a","prerequisites":[{"id":"b-6362773778e8ee0caabb0e39","url":"#b-6362773778e8ee0caabb0e39","role":"section-title"},{"id":"b-27a45f724955fef9bd94b76d","url":"#b-27a45f724955fef9bd94b76d","role":"problem-context"}],"relatedBlocks":[{"id":"b-5e05bacb7c023a91ee20a16a","url":"#b-5e05bacb7c023a91ee20a16a","relation":"same-reasoning"},{"id":"b-0fe56de7f2ef4e88eaf3ff5b","url":"#b-0fe56de7f2ef4e88eaf3ff5b","relation":"same-reasoning"},{"id":"b-47af5d3609268774a299b4ff","url":"#b-47af5d3609268774a299b4ff","relation":"same-reasoning"},{"id":"b-ea0d64617d33b325ad2fd149","url":"#b-ea0d64617d33b325ad2fd149","relation":"same-reasoning"}]},{"id":"b-5e05bacb7c023a91ee20a16a","type":"p","scope":"dense-detector-ssd-extension","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-5e05bacb7c023a91ee20a16a","content":"一阶段检测不先为每个 proposal 再做一套区域分类，而是在一个或多个特征网格上直接预测类别与位置。这个改变简化了输出路径，却没有消除参考几何、目标匹配、正负不平衡、重复预测与定位评价。","sha256":"c3fba1f619d9fa82c144218a2e5408f99396070490631920a72f7ead7b84d288","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-training-extensions","title":"选读：输出组织与训练分配还可以怎样改变？"},"page":{"id":"dense-detector-shared-problems","title":"一阶段检测简化哪一段，保留哪些问题？","number":4},"sectionTitle":"选读：一阶段检测保留了哪些相同问题","currentQuestion":"在已掌握区域、像素和实例后，按问题阅读结构与匹配扩展。","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":"dense-detector-ssd-extension","formulas":[],"previousBlock":"b-5aeffabcf350b1df8c2f3e65","nextBlock":"b-0fe56de7f2ef4e88eaf3ff5b","prerequisites":[{"id":"b-5aeffabcf350b1df8c2f3e65","url":"#b-5aeffabcf350b1df8c2f3e65","role":"section-title"},{"id":"b-27a45f724955fef9bd94b76d","url":"#b-27a45f724955fef9bd94b76d","role":"problem-context"}],"relatedBlocks":[{"id":"b-5aeffabcf350b1df8c2f3e65","url":"#b-5aeffabcf350b1df8c2f3e65","relation":"same-reasoning"},{"id":"b-0fe56de7f2ef4e88eaf3ff5b","url":"#b-0fe56de7f2ef4e88eaf3ff5b","relation":"same-reasoning"},{"id":"b-47af5d3609268774a299b4ff","url":"#b-47af5d3609268774a299b4ff","relation":"same-reasoning"},{"id":"b-ea0d64617d33b325ad2fd149","url":"#b-ea0d64617d33b325ad2fd149","relation":"same-reasoning"}]},{"id":"b-0fe56de7f2ef4e88eaf3ff5b","type":"p","scope":"dense-detector-ssd-extension","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-0fe56de7f2ef4e88eaf3ff5b","content":"若 16×16 网格每格有 4 个参考框，共 1024 项；每项预测含背景的 3 类 logits 与 4 个偏移，就有 7168 个输出数。这些数还是结构化候选，需按同一位置顺序与参考框配对，随后解码并整理。张量 reshape 顺序错误时，模型的偏移可能被贴到另一个 anchor 上。","sha256":"7d8fe2f5f510715c2b69ed15383f5797f594930db9819d67d883ae2d3c44722e","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-training-extensions","title":"选读：输出组织与训练分配还可以怎样改变？"},"page":{"id":"dense-detector-shared-problems","title":"一阶段检测简化哪一段，保留哪些问题？","number":4},"sectionTitle":"选读：一阶段检测保留了哪些相同问题","currentQuestion":"在已掌握区域、像素和实例后，按问题阅读结构与匹配扩展。","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":"dense-detector-ssd-extension","formulas":[],"previousBlock":"b-5e05bacb7c023a91ee20a16a","nextBlock":"b-47af5d3609268774a299b4ff","prerequisites":[{"id":"b-5aeffabcf350b1df8c2f3e65","url":"#b-5aeffabcf350b1df8c2f3e65","role":"section-title"},{"id":"b-27a45f724955fef9bd94b76d","url":"#b-27a45f724955fef9bd94b76d","role":"problem-context"}],"relatedBlocks":[{"id":"b-5aeffabcf350b1df8c2f3e65","url":"#b-5aeffabcf350b1df8c2f3e65","relation":"same-reasoning"},{"id":"b-5e05bacb7c023a91ee20a16a","url":"#b-5e05bacb7c023a91ee20a16a","relation":"same-reasoning"},{"id":"b-47af5d3609268774a299b4ff","url":"#b-47af5d3609268774a299b4ff","relation":"same-reasoning"},{"id":"b-ea0d64617d33b325ad2fd149","url":"#b-ea0d64617d33b325ad2fd149","relation":"same-reasoning"}]},{"id":"b-47af5d3609268774a299b4ff","type":"p","scope":"dense-detector-ssd-extension","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-47af5d3609268774a299b4ff","content":"D2L TinySSD 在多尺度特征上构造参考框与预测。当前 LAB10 为保持 CPU 可运行，用 12×12 单层网格、每格两个尺寸，共 288 项；它保留图像输入到框输出的学习链，但不等于多尺度模型的全部能力。","sha256":"159608d00faa18d3a91c3edd497cb1a181cb69965e0d2c41b1161b30140fa3ae","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-training-extensions","title":"选读：输出组织与训练分配还可以怎样改变？"},"page":{"id":"dense-detector-shared-problems","title":"一阶段检测简化哪一段，保留哪些问题？","number":4},"sectionTitle":"选读：一阶段检测保留了哪些相同问题","currentQuestion":"在已掌握区域、像素和实例后，按问题阅读结构与匹配扩展。","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":"dense-detector-ssd-extension","formulas":[],"previousBlock":"b-0fe56de7f2ef4e88eaf3ff5b","nextBlock":"b-ea0d64617d33b325ad2fd149","prerequisites":[{"id":"b-5aeffabcf350b1df8c2f3e65","url":"#b-5aeffabcf350b1df8c2f3e65","role":"section-title"},{"id":"b-27a45f724955fef9bd94b76d","url":"#b-27a45f724955fef9bd94b76d","role":"problem-context"}],"relatedBlocks":[{"id":"b-5aeffabcf350b1df8c2f3e65","url":"#b-5aeffabcf350b1df8c2f3e65","relation":"same-reasoning"},{"id":"b-5e05bacb7c023a91ee20a16a","url":"#b-5e05bacb7c023a91ee20a16a","relation":"same-reasoning"},{"id":"b-0fe56de7f2ef4e88eaf3ff5b","url":"#b-0fe56de7f2ef4e88eaf3ff5b","relation":"same-reasoning"},{"id":"b-ea0d64617d33b325ad2fd149","url":"#b-ea0d64617d33b325ad2fd149","relation":"same-reasoning"}]},{"id":"b-ea0d64617d33b325ad2fd149","type":"p","scope":"dense-detector-ssd-extension","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-ea0d64617d33b325ad2fd149","content":"本条作为从主讲移出的深入阅读：理解它有助于读实验代码，不要求课堂再开 SSD 全章。FPN 关注怎样形成不同尺度的表示，SSD 关注怎样在特征上预测；两者都出现多尺度，不意味着同一机制。","sha256":"6b8c8150ee61580fb752c1fde1e79a70c46e4b0ffcd07319b820b1e4ce5c430e","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-training-extensions","title":"选读：输出组织与训练分配还可以怎样改变？"},"page":{"id":"dense-detector-shared-problems","title":"一阶段检测简化哪一段，保留哪些问题？","number":4},"sectionTitle":"选读：一阶段检测保留了哪些相同问题","currentQuestion":"在已掌握区域、像素和实例后，按问题阅读结构与匹配扩展。","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":"dense-detector-ssd-extension","formulas":[],"previousBlock":"b-47af5d3609268774a299b4ff","nextBlock":"b-0dfd6cc71975d9fd709ae674","prerequisites":[{"id":"b-5aeffabcf350b1df8c2f3e65","url":"#b-5aeffabcf350b1df8c2f3e65","role":"section-title"},{"id":"b-27a45f724955fef9bd94b76d","url":"#b-27a45f724955fef9bd94b76d","role":"problem-context"}],"relatedBlocks":[{"id":"b-5aeffabcf350b1df8c2f3e65","url":"#b-5aeffabcf350b1df8c2f3e65","relation":"same-reasoning"},{"id":"b-5e05bacb7c023a91ee20a16a","url":"#b-5e05bacb7c023a91ee20a16a","relation":"same-reasoning"},{"id":"b-0fe56de7f2ef4e88eaf3ff5b","url":"#b-0fe56de7f2ef4e88eaf3ff5b","relation":"same-reasoning"},{"id":"b-47af5d3609268774a299b4ff","url":"#b-47af5d3609268774a299b4ff","relation":"same-reasoning"}]},{"id":"b-0dfd6cc71975d9fd709ae674","type":"h2","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-0dfd6cc71975d9fd709ae674","content":"## 沿这个问题继续","sha256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-training-extensions","title":"选读：输出组织与训练分配还可以怎样改变？"},"page":{"id":"dense-detector-shared-problems","title":"一阶段检测简化哪一段，保留哪些问题？","number":4},"sectionTitle":"一阶段检测简化哪一段，保留哪些问题？","currentQuestion":"在已掌握区域、像素和实例后，按问题阅读结构与匹配扩展。","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ea0d64617d33b325ad2fd149","nextBlock":"b-73941df9cf56bf7300cec645","prerequisites":[{"id":"b-5aeffabcf350b1df8c2f3e65","url":"#b-5aeffabcf350b1df8c2f3e65","role":"section-title"},{"id":"b-27a45f724955fef9bd94b76d","url":"#b-27a45f724955fef9bd94b76d","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-73941df9cf56bf7300cec645","type":"p","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-73941df9cf56bf7300cec645","content":"沿 CLS07 的公式定位疑问，再选 LAB09 解析练习或 LAB10 完整训练。","sha256":"a3163e9b7618b50175088fce50e8c02bf38584dbcb466052e3f6745aa91110a4","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-training-extensions","title":"选读：输出组织与训练分配还可以怎样改变？"},"page":{"id":"dense-detector-shared-problems","title":"一阶段检测简化哪一段，保留哪些问题？","number":4},"sectionTitle":"一阶段检测简化哪一段，保留哪些问题？","currentQuestion":"在已掌握区域、像素和实例后，按问题阅读结构与匹配扩展。","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-0dfd6cc71975d9fd709ae674","nextBlock":"b-bd3a0c21367ca8aa5cba72a0","prerequisites":[{"id":"b-0dfd6cc71975d9fd709ae674","url":"#b-0dfd6cc71975d9fd709ae674","role":"section-title"},{"id":"b-27a45f724955fef9bd94b76d","url":"#b-27a45f724955fef9bd94b76d","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-bd3a0c21367ca8aa5cba72a0","type":"h4","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p32.html#b-bd3a0c21367ca8aa5cba72a0","content":"#### 按问题回查的讲解与短练习（1）","sha256":"0480869f0a66009b015a757d67aed6061011e378987f98a024f4e8bfc5386152","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-training-extensions","title":"选读：输出组织与训练分配还可以怎样改变？"},"page":{"id":"dense-detector-shared-problems","title":"一阶段检测简化哪一段，保留哪些问题？","number":4},"sectionTitle":"一阶段检测简化哪一段，保留哪些问题？","currentQuestion":"在已掌握区域、像素和实例后，按问题阅读结构与匹配扩展。","sourceSlides":["l3-r2-p32","l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-73941df9cf56bf7300cec645","nextBlock":null,"prerequisites":[{"id":"b-0dfd6cc71975d9fd709ae674","url":"#b-0dfd6cc71975d9fd709ae674","role":"section-title"},{"id":"b-27a45f724955fef9bd94b76d","url":"#b-27a45f724955fef9bd94b76d","role":"problem-context"}],"relatedBlocks":[]}]}