{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"检测错在哪里，要分开看","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html","markdown":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.md","pages":[{"id":"detection-error-types","title":"在同一场景里辨认四类检测错误","unit":"检测分数很高，错误究竟在哪里？","unitId":"detection-errors-and-evidence","chapter":"II","number":1},{"id":"detection-iou-scale","title":"同样偏移一个单位，为什么小框损失更大？","unit":"检测分数很高，错误究竟在哪里？","unitId":"detection-errors-and-evidence","chapter":"II","number":2},{"id":"detection-ap-ranking","title":"沿三个预测算出 AP=5/6","unit":"检测分数很高，错误究竟在哪里？","unitId":"detection-errors-and-evidence","chapter":"II","number":3},{"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":4},{"id":"detector-nms-measured-tradeoff","title":"只改变 NMS，检查新增框究竟是什么","unit":"检测分数很高，错误究竟在哪里？","unitId":"detection-errors-and-evidence","chapter":"II","number":5},{"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":6},{"id":"focal-weighting-and-assignment","title":"把难例权重与对象匹配分成两件事","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":7},{"id":"set-matching-lab20-costs","title":"选读 LAB20：沿真实输入核对交叉匹配","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":8},{"id":"spatial-practice-selection","title":"选择一项完整实验，不按编号机械运行","unit":"带着空间接口，怎样进入实验与语言任务？","unitId":"spatial-to-language-and-practice","chapter":"II","number":9}],"aliases":[{"oldId":"explanation","targetId":"detection-error-types","operation":"split-explanation-range"},{"oldId":"b-64bf1a77b92d4081e4fc4d7c","targetId":"b-64bf1a77b92d4081e4fc4d7c","page":"detection-error-types","operation":"move","beforeSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","afterSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","reason":"Same source body in authored semantic page"},{"oldId":"b-d088b3e76f9d0388a3f7432e","targetId":"b-d088b3e76f9d0388a3f7432e","page":"detection-error-types","operation":"move","beforeSHA256":"34863663e01069ccbaf4a6a7f628dcef0684623018ecc7e3467e2d99edb025ff","afterSHA256":"34863663e01069ccbaf4a6a7f628dcef0684623018ecc7e3467e2d99edb025ff","reason":"Same source body in authored semantic page"},{"oldId":"b-b2f11c219386640541a85985","targetId":"b-b2f11c219386640541a85985","page":"detection-error-types","operation":"move","beforeSHA256":"66f71cd1faef17a7a06a83c631491bac5b4f376e49bca8b1f468f3604df1787a","afterSHA256":"66f71cd1faef17a7a06a83c631491bac5b4f376e49bca8b1f468f3604df1787a","reason":"Same source body in authored semantic page"},{"oldId":"b-8977e12ebaa6e9ff0f27a9e5","targetId":"b-8977e12ebaa6e9ff0f27a9e5","page":"detection-error-types","operation":"rewrite","beforeSHA256":"d1d18ca7a62fd72cf6ba77ec894d32b4c260e3d938c3e97629af00e1b83b6249","afterSHA256":"ed8f5ddaffec355d6f87a2b8a9b8ce4677a9d5d79628fe1221aef0f10985ce40","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-a1d097d0cb38deea92574693","targetId":"b-a1d097d0cb38deea92574693","page":"detection-error-types","operation":"move","beforeSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","afterSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","reason":"Same source body in authored semantic page"},{"oldId":"b-fc19e928830bdf086dac4c75","targetId":"b-fc19e928830bdf086dac4c75","page":"detection-error-types","operation":"move","beforeSHA256":"11b6a36079629d688f2212aef37d7b0bc1ee05d42e8fd90c010197558bdc1814","afterSHA256":"11b6a36079629d688f2212aef37d7b0bc1ee05d42e8fd90c010197558bdc1814","reason":"Same source body in authored semantic page"},{"oldId":"b-5e44ab62f89be6158a12a289","targetId":"b-5e44ab62f89be6158a12a289","page":"detection-error-types","operation":"move","beforeSHA256":"625598db0c07a9706779621404e5f6604e853fa21cf7184a73b5174519903426","afterSHA256":"625598db0c07a9706779621404e5f6604e853fa21cf7184a73b5174519903426","reason":"Same source body in authored semantic page"},{"oldId":"b-baef258fe7f0190179eb7f20","targetId":"b-baef258fe7f0190179eb7f20","page":"detection-error-types","operation":"move","beforeSHA256":"85a133647510709db321d1ab127575a3be0189ea231e185a5541203bdfe76923","afterSHA256":"85a133647510709db321d1ab127575a3be0189ea231e185a5541203bdfe76923","reason":"Same source body in authored semantic page"},{"oldId":"b-ac867b2ee05b5d4282034cf8","targetId":"b-ac867b2ee05b5d4282034cf8","page":"detection-error-types","operation":"move","beforeSHA256":"438643bfe5d3a3a2df9198ae5d16b55ef58e5a4bf998d0907008757b823e5c7e","afterSHA256":"438643bfe5d3a3a2df9198ae5d16b55ef58e5a4bf998d0907008757b823e5c7e","reason":"Same source body in authored semantic page"},{"oldId":"b-aa6a0f6865a9caed8bbc2c92","targetId":"b-aa6a0f6865a9caed8bbc2c92","page":"detection-error-types","operation":"move","beforeSHA256":"1c17a5c378a86dfadcda6212da526d1c5e42c3be7174851c650a99e2a399f718","afterSHA256":"1c17a5c378a86dfadcda6212da526d1c5e42c3be7174851c650a99e2a399f718","reason":"Same source body in authored semantic page"},{"oldId":"b-7f767f83b81f98f6c8f1679c","targetId":"b-7f767f83b81f98f6c8f1679c","page":"detection-error-types","operation":"move","beforeSHA256":"161e01cfddf4f5d3676eeeae48bf1aa5448d1bf4a937f1ed0dd2c3a505744f6c","afterSHA256":"161e01cfddf4f5d3676eeeae48bf1aa5448d1bf4a937f1ed0dd2c3a505744f6c","reason":"Same source body in authored semantic page"},{"oldId":"b-45f4cd52521fb3d2d5ca5518","targetId":"b-45f4cd52521fb3d2d5ca5518","page":"detection-error-types","operation":"move","beforeSHA256":"3c95449d64c6162fc6d26dc6b847dc41e5fb98e3a426ab73f818606e899fe970","afterSHA256":"3c95449d64c6162fc6d26dc6b847dc41e5fb98e3a426ab73f818606e899fe970","reason":"Same source body in authored semantic page"},{"oldId":"b-7751670ee8d28bd178bacb30","targetId":"b-7751670ee8d28bd178bacb30","page":"detection-error-types","operation":"move","beforeSHA256":"67909a5571cf677aa33c8bbebec12862743647fed3a514b053200dc14f9073ec","afterSHA256":"67909a5571cf677aa33c8bbebec12862743647fed3a514b053200dc14f9073ec","reason":"Same source body in authored semantic page"},{"oldId":"b-08f9196cd5af3f3a4cb7658a","targetId":"b-08f9196cd5af3f3a4cb7658a","page":"detection-error-types","operation":"move","beforeSHA256":"85bf8f8c61ed1dc61b0429b42347ee6cd22503e01ae6eb28898ae7444effc557","afterSHA256":"85bf8f8c61ed1dc61b0429b42347ee6cd22503e01ae6eb28898ae7444effc557","reason":"Same source body in authored semantic page"},{"oldId":"b-d3fd3603519dd1cd47dfc733","targetId":"b-d3fd3603519dd1cd47dfc733","page":"detection-error-types","operation":"move","beforeSHA256":"0795a1c2c3a5f46131c688973365aa9d5578d64d69115c2f2ac86abfceb9017b","afterSHA256":"0795a1c2c3a5f46131c688973365aa9d5578d64d69115c2f2ac86abfceb9017b","reason":"Same source body in authored semantic page"},{"oldId":"b-a44dc13827e80475cb334195","targetId":"b-a44dc13827e80475cb334195","page":"detection-error-types","operation":"move","beforeSHA256":"7f72fbde1c37e14bf3cbf40bfd166f8ea7b1ed5753231eadd9f708a608e2ab34","afterSHA256":"7f72fbde1c37e14bf3cbf40bfd166f8ea7b1ed5753231eadd9f708a608e2ab34","reason":"Same source body in authored semantic page"},{"oldId":"b-d755628b87c67d39e335413d","targetId":"b-d755628b87c67d39e335413d","page":"detection-error-types","operation":"move","beforeSHA256":"b97001e9a17f95cf7553d356f4edc50ba2c4bdec66f3958f36c8418e7c89a80f","afterSHA256":"b97001e9a17f95cf7553d356f4edc50ba2c4bdec66f3958f36c8418e7c89a80f","reason":"Same source body in authored semantic page"},{"oldId":"b-3cedc09498b98f6dcfb3f9de","targetId":"b-3cedc09498b98f6dcfb3f9de","page":"detection-error-types","operation":"move","beforeSHA256":"3c99309ee8c033d21924577ef0c127ef12c482975e3a962f00a9a9ddd2594817","afterSHA256":"3c99309ee8c033d21924577ef0c127ef12c482975e3a962f00a9a9ddd2594817","reason":"Same source body in authored semantic page"},{"oldId":"b-6fe0c4fd0212781e3b41229e","targetId":"b-6fe0c4fd0212781e3b41229e","page":"detection-error-types","operation":"move","beforeSHA256":"b63bae78cf2e850581b5076c1a01c03ad16c3929d8de61985157002716052bed","afterSHA256":"b63bae78cf2e850581b5076c1a01c03ad16c3929d8de61985157002716052bed","reason":"Same source body in authored semantic page"},{"oldId":"b-fb58cb8fc72a8eace6edb8dc","targetId":"b-fb58cb8fc72a8eace6edb8dc","page":"detection-error-types","operation":"move","beforeSHA256":"8af195591d37f975badc7c7752dc808cc46d486aec78e959f627d4907b525db1","afterSHA256":"8af195591d37f975badc7c7752dc808cc46d486aec78e959f627d4907b525db1","reason":"Same source body in authored semantic page"},{"oldId":"b-7d2547a32220bd364a2e82f6","targetId":"b-7d2547a32220bd364a2e82f6","page":"detection-iou-scale","operation":"move","beforeSHA256":"772bea1139ea63f9c886ce4619fee105b77b710d53ce4e8d202d7c0582a13a44","afterSHA256":"772bea1139ea63f9c886ce4619fee105b77b710d53ce4e8d202d7c0582a13a44","reason":"Same source body in authored semantic page"},{"oldId":"b-fa073fa3135ab9eb6527aa01","targetId":"b-fa073fa3135ab9eb6527aa01","page":"detection-iou-scale","operation":"move","beforeSHA256":"bf5d31cadfd50ca284a60bb547f67a52f4c270db4c300a68429ac3e2748f7219","afterSHA256":"bf5d31cadfd50ca284a60bb547f67a52f4c270db4c300a68429ac3e2748f7219","reason":"Same source body in authored semantic page"},{"oldId":"b-4d87b6a7ecf1d9018df56b02","targetId":"b-4d87b6a7ecf1d9018df56b02","page":"detection-iou-scale","operation":"move","beforeSHA256":"a4e5d6aa0c8f013ed1b565b3ac5097f4deb419989596d8302954e25684358f1e","afterSHA256":"a4e5d6aa0c8f013ed1b565b3ac5097f4deb419989596d8302954e25684358f1e","reason":"Same source body in authored semantic page"},{"oldId":"b-e4876a9e17336ce7e0160b91","targetId":"b-e4876a9e17336ce7e0160b91","page":"detection-iou-scale","operation":"move","beforeSHA256":"1ed9c49b9e5eb73a79b5965633c6d3a0f4eb6ecd61881ccb591c7de5896226b8","afterSHA256":"1ed9c49b9e5eb73a79b5965633c6d3a0f4eb6ecd61881ccb591c7de5896226b8","reason":"Same source body in authored semantic page"},{"oldId":"b-680e48950f38f34bd1867427","targetId":"b-680e48950f38f34bd1867427","page":"detection-ap-ranking","operation":"move","beforeSHA256":"bcbcb881073dbfa272a37bec76d677c7908f0adfdf2341e99ac4edd457e37673","afterSHA256":"bcbcb881073dbfa272a37bec76d677c7908f0adfdf2341e99ac4edd457e37673","reason":"Same source body in authored semantic page"},{"oldId":"b-7c5f00cfec09391321e12df1","targetId":"b-7c5f00cfec09391321e12df1","page":"detection-ap-ranking","operation":"move","beforeSHA256":"bdfffe03be3a24a07e121a89d5e62e7a9d6c0b5dfea1399e91bce266a771302c","afterSHA256":"bdfffe03be3a24a07e121a89d5e62e7a9d6c0b5dfea1399e91bce266a771302c","reason":"Same source body in authored semantic page"},{"oldId":"b-0aea64a92001f00e98b6d3f3","targetId":"b-0aea64a92001f00e98b6d3f3","page":"detection-ap-ranking","operation":"rewrite","beforeSHA256":"b922f22fe318510b5aeea72e28dba15885fa89fbf5f59a70d040b5b71e9fd3c0","afterSHA256":"81aa179742271e79995d786f39ca4f5d89b2277a9542f0c4d51f371a67dd84d4","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-5cf60d743e5672214001395b","targetId":"b-5cf60d743e5672214001395b","page":"detection-ap-ranking","operation":"move","beforeSHA256":"1abe3fad38fc2fbcdd4c804056eb02bf0d5e8a5ccc49f372f8533156f282f3d7","afterSHA256":"1abe3fad38fc2fbcdd4c804056eb02bf0d5e8a5ccc49f372f8533156f282f3d7","reason":"Same source body in authored semantic page"},{"oldId":"b-e7be383c891af82e100ff4b7","targetId":"b-e7be383c891af82e100ff4b7","page":"detection-ap-ranking","operation":"move","beforeSHA256":"ae0bbee6493290fced1bf687bb5f343334523245c57f2285e00af558b552c39f","afterSHA256":"ae0bbee6493290fced1bf687bb5f343334523245c57f2285e00af558b552c39f","reason":"Same source body in authored semantic page"},{"oldId":"b-ba994411c61ab43105b70a14","targetId":"b-ba994411c61ab43105b70a14","page":"detector-lab10-evidence","operation":"move","beforeSHA256":"76f6a9ca0e69cd8094fa23bd7ccbf42fb6bb705fb6fa0719bfe6760e58a0769f","afterSHA256":"76f6a9ca0e69cd8094fa23bd7ccbf42fb6bb705fb6fa0719bfe6760e58a0769f","reason":"Same source body in authored semantic page"},{"oldId":"b-5144d7c8819a0d53f760a354","targetId":"b-5144d7c8819a0d53f760a354","page":"detector-lab10-evidence","operation":"move","beforeSHA256":"2721b637cb6f94558f6713c359032adafb78e252bc13f52e76ce70874291e64b","afterSHA256":"2721b637cb6f94558f6713c359032adafb78e252bc13f52e76ce70874291e64b","reason":"Same source body in authored semantic page"},{"oldId":"b-eef4efc56853161c731ef41c","targetId":"b-eef4efc56853161c731ef41c","page":"detector-lab10-evidence","operation":"move","beforeSHA256":"7f64f21b960df2c59e03e9c18028330d1e9a641549b00585d29c5ca6e0f9b5b3","afterSHA256":"7f64f21b960df2c59e03e9c18028330d1e9a641549b00585d29c5ca6e0f9b5b3","reason":"Same source body in authored semantic page"},{"oldId":"b-909d02131980053ebaa30790","targetId":"b-909d02131980053ebaa30790","page":"detector-lab10-evidence","operation":"rewrite","beforeSHA256":"0ca8335cec01613a77310772ede6cd2201e74b27c01699f9c6c5125faa8c5786","afterSHA256":"ea3f2b25b776460f4e0e045b665251387ea0450a36a4ff6475b9f0188e257153","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-8797fcf5aed1bdbae216334f","targetId":"b-8797fcf5aed1bdbae216334f","page":"detector-lab10-evidence","operation":"move","beforeSHA256":"8205b30be23a99f3c1540701846c79ab37442feeb48cf7193d02963131b7f9ab","afterSHA256":"8205b30be23a99f3c1540701846c79ab37442feeb48cf7193d02963131b7f9ab","reason":"Same source body in authored semantic page"},{"oldId":"b-8a8034b89aed132313d436b1","targetId":"b-8a8034b89aed132313d436b1","page":"detector-lab10-evidence","operation":"move","beforeSHA256":"1caaa88a4dc7420a22d2566dc2b1a3e78a78af9d5f019befb84d7a50e0d3a296","afterSHA256":"1caaa88a4dc7420a22d2566dc2b1a3e78a78af9d5f019befb84d7a50e0d3a296","reason":"Same source body in authored semantic page"},{"oldId":"b-446addbdf40a6eb3aeae1349","targetId":"b-446addbdf40a6eb3aeae1349","page":"detector-nms-measured-tradeoff","operation":"move","beforeSHA256":"3264b99394a49764a79ebd0c08e1962708be37e7d92211f58fd356b70899c02f","afterSHA256":"3264b99394a49764a79ebd0c08e1962708be37e7d92211f58fd356b70899c02f","reason":"Same source body in authored semantic page"},{"oldId":"b-66e196112a2acbfcaf290c80","targetId":"b-66e196112a2acbfcaf290c80","page":"detector-nms-measured-tradeoff","operation":"move","beforeSHA256":"8b0e46c250ed7908b144b4e40151bb579c49ca95fce5168dc35905e6cd61fbbb","afterSHA256":"8b0e46c250ed7908b144b4e40151bb579c49ca95fce5168dc35905e6cd61fbbb","reason":"Same source body in authored semantic page"},{"oldId":"b-ccca693e4d6e131b7f715424","targetId":"b-ccca693e4d6e131b7f715424","page":"detector-nms-measured-tradeoff","operation":"rewrite","beforeSHA256":"40d8645433e6fab9cbb76b9a45aa3e1cb87f225d48a2be8bff29c6ed6d91df26","afterSHA256":"6568df1908194db0667c411a0d52480b9be36f1ec5c80e3aa45bba20e564fa18","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-afcfdf9f9a37423b01971fcc","targetId":"b-afcfdf9f9a37423b01971fcc","page":"detector-nms-measured-tradeoff","operation":"move","beforeSHA256":"fa71bd1d1c49b9fd714195acb6b40c28ccf5d6455b9dd9408b7c6e7101b36997","afterSHA256":"fa71bd1d1c49b9fd714195acb6b40c28ccf5d6455b9dd9408b7c6e7101b36997","reason":"Same source body in authored semantic page"},{"oldId":"b-ca9d527c53cd6cd10ea763d4","targetId":"b-ca9d527c53cd6cd10ea763d4","page":"detector-nms-measured-tradeoff","operation":"move","beforeSHA256":"07c3598d8f94057887eadde2d224c44388ad519e293cfc2ead957ce4985de933","afterSHA256":"07c3598d8f94057887eadde2d224c44388ad519e293cfc2ead957ce4985de933","reason":"Same source body in authored semantic page"},{"oldId":"b-05bd3a1b982fcbc8e92accc2","targetId":"b-05bd3a1b982fcbc8e92accc2","page":"dense-detector-shared-problems","operation":"move","beforeSHA256":"6d7a9f0134bacd810263ef2e018b0b9caefa60be8a73e1c4d75a595f843edc73","afterSHA256":"6d7a9f0134bacd810263ef2e018b0b9caefa60be8a73e1c4d75a595f843edc73","reason":"Same source body in authored semantic page"},{"oldId":"b-2cb214361a28884d64d084cc","targetId":"b-2cb214361a28884d64d084cc","page":"dense-detector-shared-problems","operation":"move","beforeSHA256":"c3fba1f619d9fa82c144218a2e5408f99396070490631920a72f7ead7b84d288","afterSHA256":"c3fba1f619d9fa82c144218a2e5408f99396070490631920a72f7ead7b84d288","reason":"Same source body in authored semantic page"},{"oldId":"b-c44aec173f967a11dbeb0c30","targetId":"b-c44aec173f967a11dbeb0c30","page":"dense-detector-shared-problems","operation":"move","beforeSHA256":"7d8fe2f5f510715c2b69ed15383f5797f594930db9819d67d883ae2d3c44722e","afterSHA256":"7d8fe2f5f510715c2b69ed15383f5797f594930db9819d67d883ae2d3c44722e","reason":"Same source body in authored semantic page"},{"oldId":"b-987676769e84a9fcbdaabe6f","targetId":"b-987676769e84a9fcbdaabe6f","page":"dense-detector-shared-problems","operation":"move","beforeSHA256":"159608d00faa18d3a91c3edd497cb1a181cb69965e0d2c41b1161b30140fa3ae","afterSHA256":"159608d00faa18d3a91c3edd497cb1a181cb69965e0d2c41b1161b30140fa3ae","reason":"Same source body in authored semantic page"},{"oldId":"b-e7412dd60f7e72eb6d7a2b81","targetId":"b-e7412dd60f7e72eb6d7a2b81","page":"dense-detector-shared-problems","operation":"move","beforeSHA256":"6b8c8150ee61580fb752c1fde1e79a70c46e4b0ffcd07319b820b1e4ce5c430e","afterSHA256":"6b8c8150ee61580fb752c1fde1e79a70c46e4b0ffcd07319b820b1e4ce5c430e","reason":"Same source body in authored semantic page"},{"oldId":"b-8bde3aebadeb8dbb78bfaa57","targetId":"b-8bde3aebadeb8dbb78bfaa57","page":"focal-weighting-and-assignment","operation":"move","beforeSHA256":"2ae031310e8b7ace430d6c798354cdf0078d2c2f282436227dab57163f170b1b","afterSHA256":"2ae031310e8b7ace430d6c798354cdf0078d2c2f282436227dab57163f170b1b","reason":"Same source body in authored semantic page"},{"oldId":"b-b0ea4c6e23604dba4fc4ba7e","targetId":"b-b0ea4c6e23604dba4fc4ba7e","page":"focal-weighting-and-assignment","operation":"move","beforeSHA256":"acad29e92db22a1ad00a309903470a4328f8c18cfdf4fac20c3446621f0f6335","afterSHA256":"acad29e92db22a1ad00a309903470a4328f8c18cfdf4fac20c3446621f0f6335","reason":"Same source body in authored semantic page"},{"oldId":"b-4120daf249592055a1d25758","targetId":"b-4120daf249592055a1d25758","page":"focal-weighting-and-assignment","operation":"move","beforeSHA256":"6d41a42fe14ef8b81dbbf4c540876335e4e04e4ecf75f45c1c416ef808b7f693","afterSHA256":"6d41a42fe14ef8b81dbbf4c540876335e4e04e4ecf75f45c1c416ef808b7f693","reason":"Same source body in authored semantic page"},{"oldId":"b-263e596b99b06ddd6280e3a2","targetId":"b-263e596b99b06ddd6280e3a2","page":"focal-weighting-and-assignment","operation":"move","beforeSHA256":"09c26eb8dd7894d9423b7b2465780fda6811ebe147e75d5d917afc5f99951c94","afterSHA256":"09c26eb8dd7894d9423b7b2465780fda6811ebe147e75d5d917afc5f99951c94","reason":"Same source body in authored semantic page"},{"oldId":"b-07496f85b77ac1623307e2eb","targetId":"b-07496f85b77ac1623307e2eb","page":"set-matching-lab20-costs","operation":"move","beforeSHA256":"e22ba271dbf8de903d354b548af45e31549397c87cf88aa2ff55aab69ec47377","afterSHA256":"e22ba271dbf8de903d354b548af45e31549397c87cf88aa2ff55aab69ec47377","reason":"Same source body in authored semantic page"},{"oldId":"b-0f92d758c53ae5d4a6f20de3","targetId":"b-0f92d758c53ae5d4a6f20de3","page":"set-matching-lab20-costs","operation":"move","beforeSHA256":"8c4a4fcaec0e6c0eb856799d47828d49dcd70cbeddd2d724b59395207d95348e","afterSHA256":"8c4a4fcaec0e6c0eb856799d47828d49dcd70cbeddd2d724b59395207d95348e","reason":"Same source body in authored semantic page"},{"oldId":"b-91a5f7f6a09aaa85af8addb8","targetId":"b-91a5f7f6a09aaa85af8addb8","page":"set-matching-lab20-costs","operation":"move","beforeSHA256":"666afb23ae32517dfc3373217aa07120f72ac5e338277cf55cef59ec35e29c96","afterSHA256":"666afb23ae32517dfc3373217aa07120f72ac5e338277cf55cef59ec35e29c96","reason":"Same source body in authored semantic page"},{"oldId":"b-efb5f1232e4b79525f38a178","targetId":"b-efb5f1232e4b79525f38a178","page":"set-matching-lab20-costs","operation":"move","beforeSHA256":"00cdf3ea105cff573d3b5e30cd194b7c8bb22cc8b37fda235912d4f9ef325925","afterSHA256":"00cdf3ea105cff573d3b5e30cd194b7c8bb22cc8b37fda235912d4f9ef325925","reason":"Same source body in authored semantic page"},{"oldId":"b-2d6fed637b1b9ac05025df8c","targetId":"b-2d6fed637b1b9ac05025df8c","page":"set-matching-lab20-costs","operation":"move","beforeSHA256":"2c901bf98cd806f477e314f0897e9351b3172089a6eab29d4f2569837c1837e6","afterSHA256":"2c901bf98cd806f477e314f0897e9351b3172089a6eab29d4f2569837c1837e6","reason":"Same source body in authored semantic page"},{"oldId":"b-d76aead57f8f31763bbf38dd","targetId":"b-d76aead57f8f31763bbf38dd","page":"spatial-practice-selection","operation":"move","beforeSHA256":"9c1d4e7b42a1c13bd684104ac7d8b248ab71cdd076f2976b642e9cd4d88212ac","afterSHA256":"9c1d4e7b42a1c13bd684104ac7d8b248ab71cdd076f2976b642e9cd4d88212ac","reason":"Same source body in authored semantic page"},{"oldId":"b-3de39f5d21e18cda37f10dc2","targetId":"b-3de39f5d21e18cda37f10dc2","page":"spatial-practice-selection","operation":"move","beforeSHA256":"0a2ecad891c31a62ee9b02e86cdba8178bd3d109c92dba9a0311468b72ee9cd0","afterSHA256":"0a2ecad891c31a62ee9b02e86cdba8178bd3d109c92dba9a0311468b72ee9cd0","reason":"Same source body in authored semantic page"},{"oldId":"b-3de7cc46ed6aec104caec368","targetId":"b-3de7cc46ed6aec104caec368","page":"spatial-practice-selection","operation":"move","beforeSHA256":"6c778237b86f2e11d4c30d82b852d6200f4b10ccec5ed4c1beb3aad463157942","afterSHA256":"6c778237b86f2e11d4c30d82b852d6200f4b10ccec5ed4c1beb3aad463157942","reason":"Same source body in authored semantic page"},{"oldId":"b-b088dad2685cf2eec1252f6f","targetId":"b-b088dad2685cf2eec1252f6f","page":"spatial-practice-selection","operation":"move","beforeSHA256":"5bd92268b823ca1189663b03a9aab5ad528b77f1824da7041e68e16660f8a280","afterSHA256":"5bd92268b823ca1189663b03a9aab5ad528b77f1824da7041e68e16660f8a280","reason":"Same source body in authored semantic page"},{"oldId":"b-c1e3af93b0c02bc155ada275","targetId":"b-c1e3af93b0c02bc155ada275","page":"spatial-practice-selection","operation":"move","beforeSHA256":"7dc68c2032c902245b20b5e27e414e7e048471fa6a74d58eab62ea8327fd9a26","afterSHA256":"7dc68c2032c902245b20b5e27e414e7e048471fa6a74d58eab62ea8327fd9a26","reason":"Same source body in authored semantic page"},{"oldId":"b-30c7edf0ddb51751ac361073","targetId":"b-30c7edf0ddb51751ac361073","page":"spatial-practice-selection","operation":"move","beforeSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","afterSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","reason":"Same source body in authored semantic page"},{"oldId":"b-a0cf6fcaa14744440d0133e4","targetId":"b-a0cf6fcaa14744440d0133e4","page":"spatial-practice-selection","operation":"move","beforeSHA256":"8f2c63167e56ee768534601ac56a55e07567c8792a5deb33e7da197b37f632c9","afterSHA256":"8f2c63167e56ee768534601ac56a55e07567c8792a5deb33e7da197b37f632c9","reason":"Same source body in authored semantic page"},{"oldId":"detection-error-types-title","targetId":"b-72ffaadc742c55db37fff809","operation":"retain-source-identity"},{"oldId":"detection-iou-scale-title","targetId":"b-de2a3476f858ddc25ae6be4d","operation":"retain-source-identity"},{"oldId":"detection-iou-scale-bridge","targetId":"b-23338d84ff8b6b413d733655","operation":"retain-source-identity"},{"oldId":"detection-ap-ranking-title","targetId":"b-1881a18f1bdac2b4584025ac","operation":"retain-source-identity"},{"oldId":"detection-ap-ranking-bridge","targetId":"b-6280ccf4cb24a8b3b9baee5a","operation":"retain-source-identity"},{"oldId":"detector-lab10-evidence-title","targetId":"b-dbe527013e51f33f30bf87bc","operation":"retain-source-identity"},{"oldId":"detector-lab10-evidence-bridge","targetId":"b-50efb17832df14fd94de7726","operation":"retain-source-identity"},{"oldId":"lab10-saved-failures","targetId":"b-f39861246e1e47fe535429ed","operation":"retain-source-identity"},{"oldId":"detector-nms-measured-tradeoff-title","targetId":"b-752f624db54b5c1ba4df738e","operation":"retain-source-identity"},{"oldId":"detector-nms-measured-tradeoff-bridge","targetId":"b-734e68dc13afaf39bbba5d88","operation":"retain-source-identity"},{"oldId":"dense-detector-shared-problems-title","targetId":"b-2bded4a5236159ca45bd550b","operation":"retain-source-identity"},{"oldId":"dense-detector-shared-problems-bridge","targetId":"b-80d1985c1080b8f662880e27","operation":"retain-source-identity"},{"oldId":"focal-weighting-and-assignment-title","targetId":"b-bc9124cedc5b09deedc151cd","operation":"retain-source-identity"},{"oldId":"focal-weighting-and-assignment-bridge","targetId":"b-079b27913a200169a4693cf2","operation":"retain-source-identity"},{"oldId":"set-matching-lab20-costs-title","targetId":"b-b7e45148839ce1d171344d73","operation":"retain-source-identity"},{"oldId":"set-matching-lab20-costs-bridge","targetId":"b-e637ab72577ef03dfac4f03d","operation":"retain-source-identity"},{"oldId":"spatial-practice-selection-title","targetId":"b-e3b9d4c1a5382a7839d845b5","operation":"retain-source-identity"},{"oldId":"spatial-practice-selection-bridge","targetId":"b-9f21524d34ae60bba14da238","operation":"retain-source-identity"}],"blocks":[{"id":"b-72ffaadc742c55db37fff809","type":"h2","scope":"detection-error-types","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-72ffaadc742c55db37fff809","content":"## 在同一场景里辨认四类检测错误","sha256":"d2b83cf3df1153be1ff6557ecdd50a0199b6bfa99302f856e94bf5625654cec0","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-64bf1a77b92d4081e4fc4d7c","prerequisites":[{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-64bf1a77b92d4081e4fc4d7c","type":"h2","scope":"explanation","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-64bf1a77b92d4081e4fc4d7c","content":"## 解释、算例与证据","sha256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-72ffaadc742c55db37fff809","nextBlock":"b-d088b3e76f9d0388a3f7432e","prerequisites":[{"id":"b-72ffaadc742c55db37fff809","url":"#b-72ffaadc742c55db37fff809","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-d088b3e76f9d0388a3f7432e","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-d088b3e76f9d0388a3f7432e","content":"# P35 · 检测错在哪里，要分开看","sha256":"34863663e01069ccbaf4a6a7f628dcef0684623018ecc7e3467e2d99edb025ff","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-64bf1a77b92d4081e4fc4d7c","nextBlock":"b-b2f11c219386640541a85985","prerequisites":[{"id":"b-64bf1a77b92d4081e4fc4d7c","url":"#b-64bf1a77b92d4081e4fc4d7c","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-b2f11c219386640541a85985","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-b2f11c219386640541a85985","content":"从错类、漏检、框偏和重复四项错误读结果。IoU 把位置偏差变成可比较几何量，AP 再加入分数排序与对象唯一匹配。沿下方积分例走完分数排序、精确率与召回率计算，再用 LAB10 的结果辨认这些错误。","sha256":"66f71cd1faef17a7a06a83c631491bac5b4f376e49bca8b1f468f3604df1787a","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d088b3e76f9d0388a3f7432e","nextBlock":"b-8977e12ebaa6e9ff0f27a9e5","prerequisites":[{"id":"b-d088b3e76f9d0388a3f7432e","url":"#b-d088b3e76f9d0388a3f7432e","role":"section-title"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-8977e12ebaa6e9ff0f27a9e5","type":"figure","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-8977e12ebaa6e9ff0f27a9e5","content":"图内标签与数值：Lecture 03 课堂图解 P35；漏检；没有报告蓝杯；错类别；红杯 → “花瓶”；偏位；重复；面积各 4 · 交集 2；IoU = 2 / (4 + 4 − 2) = 1/3；类别分数高，不会自动修正位置；检测错误要分开检查；完整匹配、阈值和 AP 协议在补充阅读。\n漏检\n图内标签与数值：\n真值对象没有得到匹配的预测，召回会降低。错类别\n图内标签与数值：\n框覆盖对象，但类别不符，不能当作正确检测。框偏位\n图内标签与数值：\n类别对了，位置仍可能过不了 IoU 阈值。重复预测\n图内标签与数值：\n一个真值最多匹配一次；后来的重复框计作假阳性。\nIoU 算例：两个框面积各为4，交集为2；并集=4+4−2=6，IoU=2/6=1/3。类别分数高不会自动修正位置。\n同一场景的四种检测错误，以及面积各 4、交集 2 的框对，算得 IoU=1/3。 [查看对应图解状态](/#/l3-r2-p35/0)[查看完整图解](/course/figures/l3-r2-p35-0.html)","sha256":"ed8f5ddaffec355d6f87a2b8a9b8ce4677a9d5d79628fe1221aef0f10985ce40","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"同一场景的四种检测错误，以及面积各 4、交集 2 的框对，算得 IoU=1/3。 查看对应图解状态","labels":["漏检","没有报告蓝杯","错类别","红杯 → “花瓶”","偏位","重复","面积各 4 · 交集 2","IoU = 2 / (4 + 4 − 2) = 1/3","类别分数高，不会自动修正位置","检测错误要分开检查；完整匹配、阈值和 AP 协议在补充阅读。"],"description":"漏检真值对象没有得到匹配的预测，召回会降低。错类别框覆盖对象，但类别不符，不能当作正确检测。框偏位类别对了，位置仍可能过不了 IoU 阈值。重复预测一个真值最多匹配一次；后来的重复框计作假阳性。IoU 算例：两个框面积各为4，交集为2；并集=4+4−2=6，IoU=2/6=1/3。类别分数高不会自动修正位置。","evidenceIdentity":"同一场景的四种检测错误，以及面积各 4、交集 2 的框对，算得 IoU=1/3。 查看对应图解状态","images":[{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""},{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""},{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""},{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""},{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""},{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""},{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""},{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""}]},"previousBlock":"b-b2f11c219386640541a85985","nextBlock":"b-a1d097d0cb38deea92574693","prerequisites":[{"id":"b-d088b3e76f9d0388a3f7432e","url":"#b-d088b3e76f9d0388a3f7432e","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-a1d097d0cb38deea92574693","type":"h2","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-a1d097d0cb38deea92574693","content":"## 课堂讲述","sha256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-8977e12ebaa6e9ff0f27a9e5","nextBlock":"b-fc19e928830bdf086dac4c75","prerequisites":[{"id":"b-d088b3e76f9d0388a3f7432e","url":"#b-d088b3e76f9d0388a3f7432e","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-fc19e928830bdf086dac4c75","type":"p","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-fc19e928830bdf086dac4c75","content":"现在检验系统，不能只问它有没有认出“杯子”。这个结果漏了一只，那个类别错了，这个位置偏了，另一个把同一只报了两次。这些错误需要分别观察。\n位置常用交并比描述：相交面积除以覆盖的总面积。例子中两个框面积都是四，交集为二，交并比就是三分之一。类别分数高，不会自动把这个位置变得准确。\n完整评价还要说明预测怎样与真值匹配，怎样处理重复和阈值。今天不手推整套 AP，但要留下这项判断：当你说某个预训练表示改善了检测，必须知道改善体现在哪种任务错误，而不只看一列总分。这也为接下来的像素任务准备了不同的检查方式。","sha256":"11b6a36079629d688f2212aef37d7b0bc1ee05d42e8fd90c010197558bdc1814","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-a1d097d0cb38deea92574693","nextBlock":"b-f4c7f33ac45a9db237038e32","prerequisites":[{"id":"b-a1d097d0cb38deea92574693","url":"#b-a1d097d0cb38deea92574693","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-f4c7f33ac45a9db237038e32","type":"h4","scope":"detection-error-types","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-f4c7f33ac45a9db237038e32","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":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-fc19e928830bdf086dac4c75","nextBlock":"b-5e44ab62f89be6158a12a289","prerequisites":[{"id":"b-a1d097d0cb38deea92574693","url":"#b-a1d097d0cb38deea92574693","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-5e44ab62f89be6158a12a289","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-5e44ab62f89be6158a12a289","content":"IoU 面积算例使用两个 2×2 连续方框，横向错开1，交集2、并集6，IoU=1/3。四种错误是教学构造，不是本轮推理。","sha256":"625598db0c07a9706779621404e5f6604e853fa21cf7184a73b5174519903426","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f4c7f33ac45a9db237038e32","nextBlock":"b-1023d9b135661d96377aed43","prerequisites":[{"id":"b-f4c7f33ac45a9db237038e32","url":"#b-f4c7f33ac45a9db237038e32","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-1023d9b135661d96377aed43","type":"h4","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-1023d9b135661d96377aed43","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":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-5e44ab62f89be6158a12a289","nextBlock":"b-baef258fe7f0190179eb7f20","prerequisites":[{"id":"b-f4c7f33ac45a9db237038e32","url":"#b-f4c7f33ac45a9db237038e32","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-baef258fe7f0190179eb7f20","type":"ul","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-baef258fe7f0190179eb7f20","content":"- [检测错误要经过排序与匹配｜问题从哪里出现](/supplement.html?return=l3-r2-p35/0#/l3-detection-evaluation-problem/0) 本页说明\n- [检测错误要经过排序与匹配｜核心改动怎样工作](/supplement.html?return=l3-r2-p35/0#/l3-detection-evaluation-mechanism/0) 本页说明\n- [检测错误要经过排序与匹配｜一起算一个完整例子](/supplement.html?return=l3-r2-p35/0#/l3-detection-evaluation-example/0) 本页说明\n- [检测错误要经过排序与匹配｜怎样检查这项解释](/supplement.html?return=l3-r2-p35/0#/l3-detection-evaluation-evidence/0) 本页说明\n- [检测错误要经过排序与匹配｜什么结论还不能推出](/supplement.html?return=l3-r2-p35/0#/l3-detection-evaluation-boundary/0) 本页说明\n- [一阶段检测把预测铺满特征图｜怎样检查这项解释](/supplement.html?return=l3-r2-p35/0#/l3-dense-one-stage-evidence/0) L03补充阅读\n- [一阶段检测把预测铺满特征图｜什么结论还不能推出](/supplement.html?return=l3-r2-p35/0#/l3-dense-one-stage-boundary/0) L03补充阅读\n- [类别不平衡与集合输出是两种问题｜问题从哪里出现](/supplement.html?return=l3-r2-p35/0#/l3-focal-set-problem/0) L03补充阅读\n- [类别不平衡与集合输出是两种问题｜核心改动怎样工作](/supplement.html?return=l3-r2-p35/0#/l3-focal-set-mechanism/0) L03补充阅读\n- [类别不平衡与集合输出是两种问题｜一起算一个完整例子](/supplement.html?return=l3-r2-p35/0#/l3-focal-set-example/0) L03补充阅读\n- [类别不平衡与集合输出是两种问题｜怎样检查这项解释](/supplement.html?return=l3-r2-p35/0#/l3-focal-set-evidence/0) L03补充阅读\n- [类别不平衡与集合输出是两种问题｜什么结论还不能推出](/supplement.html?return=l3-r2-p35/0#/l3-focal-set-boundary/0) L03补充阅读","sha256":"85a133647510709db321d1ab127575a3be0189ea231e185a5541203bdfe76923","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-1023d9b135661d96377aed43","nextBlock":"b-ac867b2ee05b5d4282034cf8","prerequisites":[{"id":"b-1023d9b135661d96377aed43","url":"#b-1023d9b135661d96377aed43","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-ac867b2ee05b5d4282034cf8","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-ac867b2ee05b5d4282034cf8","content":"- [检测错误要经过排序与匹配｜问题从哪里出现](/supplement.html?return=l3-r2-p35/0#/l3-detection-evaluation-problem/0) 本页说明","sha256":"438643bfe5d3a3a2df9198ae5d16b55ef58e5a4bf998d0907008757b823e5c7e","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-baef258fe7f0190179eb7f20","nextBlock":"b-aa6a0f6865a9caed8bbc2c92","prerequisites":[{"id":"b-1023d9b135661d96377aed43","url":"#b-1023d9b135661d96377aed43","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-aa6a0f6865a9caed8bbc2c92","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-aa6a0f6865a9caed8bbc2c92","content":"- [检测错误要经过排序与匹配｜核心改动怎样工作](/supplement.html?return=l3-r2-p35/0#/l3-detection-evaluation-mechanism/0) 本页说明","sha256":"1c17a5c378a86dfadcda6212da526d1c5e42c3be7174851c650a99e2a399f718","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ac867b2ee05b5d4282034cf8","nextBlock":"b-7f767f83b81f98f6c8f1679c","prerequisites":[{"id":"b-1023d9b135661d96377aed43","url":"#b-1023d9b135661d96377aed43","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-7f767f83b81f98f6c8f1679c","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-7f767f83b81f98f6c8f1679c","content":"- [检测错误要经过排序与匹配｜一起算一个完整例子](/supplement.html?return=l3-r2-p35/0#/l3-detection-evaluation-example/0) 本页说明","sha256":"161e01cfddf4f5d3676eeeae48bf1aa5448d1bf4a937f1ed0dd2c3a505744f6c","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-aa6a0f6865a9caed8bbc2c92","nextBlock":"b-45f4cd52521fb3d2d5ca5518","prerequisites":[{"id":"b-1023d9b135661d96377aed43","url":"#b-1023d9b135661d96377aed43","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-45f4cd52521fb3d2d5ca5518","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-45f4cd52521fb3d2d5ca5518","content":"- [检测错误要经过排序与匹配｜怎样检查这项解释](/supplement.html?return=l3-r2-p35/0#/l3-detection-evaluation-evidence/0) 本页说明","sha256":"3c95449d64c6162fc6d26dc6b847dc41e5fb98e3a426ab73f818606e899fe970","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7f767f83b81f98f6c8f1679c","nextBlock":"b-7751670ee8d28bd178bacb30","prerequisites":[{"id":"b-1023d9b135661d96377aed43","url":"#b-1023d9b135661d96377aed43","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-7751670ee8d28bd178bacb30","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-7751670ee8d28bd178bacb30","content":"- [检测错误要经过排序与匹配｜什么结论还不能推出](/supplement.html?return=l3-r2-p35/0#/l3-detection-evaluation-boundary/0) 本页说明","sha256":"67909a5571cf677aa33c8bbebec12862743647fed3a514b053200dc14f9073ec","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-45f4cd52521fb3d2d5ca5518","nextBlock":"b-08f9196cd5af3f3a4cb7658a","prerequisites":[{"id":"b-1023d9b135661d96377aed43","url":"#b-1023d9b135661d96377aed43","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-08f9196cd5af3f3a4cb7658a","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-08f9196cd5af3f3a4cb7658a","content":"- [一阶段检测把预测铺满特征图｜怎样检查这项解释](/supplement.html?return=l3-r2-p35/0#/l3-dense-one-stage-evidence/0) L03补充阅读","sha256":"85bf8f8c61ed1dc61b0429b42347ee6cd22503e01ae6eb28898ae7444effc557","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7751670ee8d28bd178bacb30","nextBlock":"b-d3fd3603519dd1cd47dfc733","prerequisites":[{"id":"b-1023d9b135661d96377aed43","url":"#b-1023d9b135661d96377aed43","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-d3fd3603519dd1cd47dfc733","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-d3fd3603519dd1cd47dfc733","content":"- [一阶段检测把预测铺满特征图｜什么结论还不能推出](/supplement.html?return=l3-r2-p35/0#/l3-dense-one-stage-boundary/0) L03补充阅读","sha256":"0795a1c2c3a5f46131c688973365aa9d5578d64d69115c2f2ac86abfceb9017b","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-08f9196cd5af3f3a4cb7658a","nextBlock":"b-a44dc13827e80475cb334195","prerequisites":[{"id":"b-1023d9b135661d96377aed43","url":"#b-1023d9b135661d96377aed43","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-a44dc13827e80475cb334195","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-a44dc13827e80475cb334195","content":"- [类别不平衡与集合输出是两种问题｜问题从哪里出现](/supplement.html?return=l3-r2-p35/0#/l3-focal-set-problem/0) L03补充阅读","sha256":"7f72fbde1c37e14bf3cbf40bfd166f8ea7b1ed5753231eadd9f708a608e2ab34","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d3fd3603519dd1cd47dfc733","nextBlock":"b-d755628b87c67d39e335413d","prerequisites":[{"id":"b-1023d9b135661d96377aed43","url":"#b-1023d9b135661d96377aed43","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-d755628b87c67d39e335413d","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-d755628b87c67d39e335413d","content":"- [类别不平衡与集合输出是两种问题｜核心改动怎样工作](/supplement.html?return=l3-r2-p35/0#/l3-focal-set-mechanism/0) L03补充阅读","sha256":"b97001e9a17f95cf7553d356f4edc50ba2c4bdec66f3958f36c8418e7c89a80f","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-a44dc13827e80475cb334195","nextBlock":"b-3cedc09498b98f6dcfb3f9de","prerequisites":[{"id":"b-1023d9b135661d96377aed43","url":"#b-1023d9b135661d96377aed43","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-3cedc09498b98f6dcfb3f9de","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-3cedc09498b98f6dcfb3f9de","content":"- [类别不平衡与集合输出是两种问题｜一起算一个完整例子](/supplement.html?return=l3-r2-p35/0#/l3-focal-set-example/0) L03补充阅读","sha256":"3c99309ee8c033d21924577ef0c127ef12c482975e3a962f00a9a9ddd2594817","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d755628b87c67d39e335413d","nextBlock":"b-6fe0c4fd0212781e3b41229e","prerequisites":[{"id":"b-1023d9b135661d96377aed43","url":"#b-1023d9b135661d96377aed43","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-6fe0c4fd0212781e3b41229e","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-6fe0c4fd0212781e3b41229e","content":"- [类别不平衡与集合输出是两种问题｜怎样检查这项解释](/supplement.html?return=l3-r2-p35/0#/l3-focal-set-evidence/0) L03补充阅读","sha256":"b63bae78cf2e850581b5076c1a01c03ad16c3929d8de61985157002716052bed","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3cedc09498b98f6dcfb3f9de","nextBlock":"b-fb58cb8fc72a8eace6edb8dc","prerequisites":[{"id":"b-1023d9b135661d96377aed43","url":"#b-1023d9b135661d96377aed43","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-fb58cb8fc72a8eace6edb8dc","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-fb58cb8fc72a8eace6edb8dc","content":"- [类别不平衡与集合输出是两种问题｜什么结论还不能推出](/supplement.html?return=l3-r2-p35/0#/l3-focal-set-boundary/0) L03补充阅读","sha256":"8af195591d37f975badc7c7752dc808cc46d486aec78e959f627d4907b525db1","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":"detection-error-types","title":"在同一场景里辨认四类检测错误","number":1},"sectionTitle":"在同一场景里辨认四类检测错误","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6fe0c4fd0212781e3b41229e","nextBlock":null,"prerequisites":[{"id":"b-1023d9b135661d96377aed43","url":"#b-1023d9b135661d96377aed43","role":"section-title"},{"id":"b-b2f11c219386640541a85985","url":"#b-b2f11c219386640541a85985","role":"problem-context"}],"relatedBlocks":[{"id":"b-8977e12ebaa6e9ff0f27a9e5","url":"#b-8977e12ebaa6e9ff0f27a9e5","relation":"figure-and-explanation"}]},{"id":"b-de2a3476f858ddc25ae6be4d","type":"h2","scope":"detection-iou-scale","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-de2a3476f858ddc25ae6be4d","content":"## 同样偏移一个单位，为什么小框损失更大？","sha256":"2c3cebf9ba1b309d2dfb07f620800b2f4a346c9c1efea4a3d1ab051aeb957f5e","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":"detection-iou-scale","title":"同样偏移一个单位，为什么小框损失更大？","number":2},"sectionTitle":"同样偏移一个单位，为什么小框损失更大？","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-23338d84ff8b6b413d733655","prerequisites":[{"id":"b-23338d84ff8b6b413d733655","url":"#b-23338d84ff8b6b413d733655","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-23338d84ff8b6b413d733655","type":"p","scope":"detection-iou-scale","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-23338d84ff8b6b413d733655","content":"从框偏这一类错误开始，把交集、并集和对象尺度放进同一计算。","sha256":"b7f4ac85ddbf8db1ae2cb1d0c3d626b41119834b0b760c24e9da3af574dbeeae","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":"detection-iou-scale","title":"同样偏移一个单位，为什么小框损失更大？","number":2},"sectionTitle":"同样偏移一个单位，为什么小框损失更大？","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-de2a3476f858ddc25ae6be4d","nextBlock":"b-7d2547a32220bd364a2e82f6","prerequisites":[{"id":"b-de2a3476f858ddc25ae6be4d","url":"#b-de2a3476f858ddc25ae6be4d","role":"section-title"}],"relatedBlocks":[]},{"id":"b-7d2547a32220bd364a2e82f6","type":"h3","scope":"iou-and-small-boxes","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-7d2547a32220bd364a2e82f6","content":"### 同样移动一个单位，小框为什么更吃亏","sha256":"772bea1139ea63f9c886ce4619fee105b77b710d53ce4e8d202d7c0582a13a44","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":"detection-iou-scale","title":"同样偏移一个单位，为什么小框损失更大？","number":2},"sectionTitle":"同样移动一个单位，小框为什么更吃亏","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":"iou-and-small-boxes","formulas":[],"previousBlock":"b-23338d84ff8b6b413d733655","nextBlock":"b-fa073fa3135ab9eb6527aa01","prerequisites":[{"id":"b-de2a3476f858ddc25ae6be4d","url":"#b-de2a3476f858ddc25ae6be4d","role":"section-title"},{"id":"b-23338d84ff8b6b413d733655","url":"#b-23338d84ff8b6b413d733655","role":"problem-context"}],"relatedBlocks":[{"id":"b-fa073fa3135ab9eb6527aa01","url":"#b-fa073fa3135ab9eb6527aa01","relation":"same-reasoning"},{"id":"b-4d87b6a7ecf1d9018df56b02","url":"#b-4d87b6a7ecf1d9018df56b02","relation":"same-reasoning"},{"id":"b-e4876a9e17336ce7e0160b91","url":"#b-e4876a9e17336ce7e0160b91","relation":"same-reasoning"}]},{"id":"b-fa073fa3135ab9eb6527aa01","type":"p","scope":"iou-and-small-boxes","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-fa073fa3135ab9eb6527aa01","content":"两个面积均为 4 的框，交集为 2，并集为 $4+4-2=6$，所以 IoU=1/3，而不是 1/2。分母必须是覆盖总面积，不能只用真值面积。连续坐标的宽写作 $x_2-x_1$，不要混入“整数像素含两端”的加一规则。","sha256":"bf5d31cadfd50ca284a60bb547f67a52f4c270db4c300a68429ac3e2748f7219","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":"detection-iou-scale","title":"同样偏移一个单位，为什么小框损失更大？","number":2},"sectionTitle":"同样移动一个单位，小框为什么更吃亏","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":"iou-and-small-boxes","formulas":[{"latex":"4+4-2=6","definitionContext":"两个面积均为 4 的框，交集为 2，并集为 $4+4-2=6$，所以 IoU=1/3，而不是 1/2。分母必须是覆盖总面积，不能只用真值面积。连续坐标的宽写作 $x_2-x_1$，不要混入“整数像素含两端”的加一规则。\n把两个同样大小的正方形横向错开 d，边长为 w 且 $0\\le d<w$。交集面积 $(w-d)w$，并集 $(w+d)w$，于是 IoU=$(w-d)/(w+d)$。错开一个单位时，边长 10 的框 IoU=9/11，边长 2 的框却只有 1/3。绝对坐标误差相同，几何质量不同。\n检测检查还要分清错类、漏检、框偏和重复。分数只说明模型的相对确信程度，不能直接把偏掉的框变准。若同一模型 AP50 很好而 AP75 明显下降，说明它能覆盖不少对象，但更严格位置要求暴露了问题；还应查看具体预测，才能知道误差来自偏移、大小还是同类拥挤。"},{"latex":"x_2-x_1","definitionContext":"两个面积均为 4 的框，交集为 2，并集为 $4+4-2=6$，所以 IoU=1/3，而不是 1/2。分母必须是覆盖总面积，不能只用真值面积。连续坐标的宽写作 $x_2-x_1$，不要混入“整数像素含两端”的加一规则。\n把两个同样大小的正方形横向错开 d，边长为 w 且 $0\\le d<w$。交集面积 $(w-d)w$，并集 $(w+d)w$，于是 IoU=$(w-d)/(w+d)$。错开一个单位时，边长 10 的框 IoU=9/11，边长 2 的框却只有 1/3。绝对坐标误差相同，几何质量不同。\n检测检查还要分清错类、漏检、框偏和重复。分数只说明模型的相对确信程度，不能直接把偏掉的框变准。若同一模型 AP50 很好而 AP75 明显下降，说明它能覆盖不少对象，但更严格位置要求暴露了问题；还应查看具体预测，才能知道误差来自偏移、大小还是同类拥挤。"}],"previousBlock":"b-7d2547a32220bd364a2e82f6","nextBlock":"b-4d87b6a7ecf1d9018df56b02","prerequisites":[{"id":"b-7d2547a32220bd364a2e82f6","url":"#b-7d2547a32220bd364a2e82f6","role":"section-title"},{"id":"b-23338d84ff8b6b413d733655","url":"#b-23338d84ff8b6b413d733655","role":"problem-context"}],"relatedBlocks":[{"id":"b-7d2547a32220bd364a2e82f6","url":"#b-7d2547a32220bd364a2e82f6","relation":"same-reasoning"},{"id":"b-4d87b6a7ecf1d9018df56b02","url":"#b-4d87b6a7ecf1d9018df56b02","relation":"same-reasoning"},{"id":"b-e4876a9e17336ce7e0160b91","url":"#b-e4876a9e17336ce7e0160b91","relation":"same-reasoning"}]},{"id":"b-4d87b6a7ecf1d9018df56b02","type":"p","scope":"iou-and-small-boxes","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-4d87b6a7ecf1d9018df56b02","content":"把两个同样大小的正方形横向错开 d，边长为 w 且 $0\\le d<w$。交集面积 $(w-d)w$，并集 $(w+d)w$，于是 IoU=$(w-d)/(w+d)$。错开一个单位时，边长 10 的框 IoU=9/11，边长 2 的框却只有 1/3。绝对坐标误差相同，几何质量不同。","sha256":"a4e5d6aa0c8f013ed1b565b3ac5097f4deb419989596d8302954e25684358f1e","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":"detection-iou-scale","title":"同样偏移一个单位，为什么小框损失更大？","number":2},"sectionTitle":"同样移动一个单位，小框为什么更吃亏","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":"iou-and-small-boxes","formulas":[{"latex":"0\\le d<w","definitionContext":"两个面积均为 4 的框，交集为 2，并集为 $4+4-2=6$，所以 IoU=1/3，而不是 1/2。分母必须是覆盖总面积，不能只用真值面积。连续坐标的宽写作 $x_2-x_1$，不要混入“整数像素含两端”的加一规则。\n把两个同样大小的正方形横向错开 d，边长为 w 且 $0\\le d<w$。交集面积 $(w-d)w$，并集 $(w+d)w$，于是 IoU=$(w-d)/(w+d)$。错开一个单位时，边长 10 的框 IoU=9/11，边长 2 的框却只有 1/3。绝对坐标误差相同，几何质量不同。\n检测检查还要分清错类、漏检、框偏和重复。分数只说明模型的相对确信程度，不能直接把偏掉的框变准。若同一模型 AP50 很好而 AP75 明显下降，说明它能覆盖不少对象，但更严格位置要求暴露了问题；还应查看具体预测，才能知道误差来自偏移、大小还是同类拥挤。"},{"latex":"(w-d)w","definitionContext":"两个面积均为 4 的框，交集为 2，并集为 $4+4-2=6$，所以 IoU=1/3，而不是 1/2。分母必须是覆盖总面积，不能只用真值面积。连续坐标的宽写作 $x_2-x_1$，不要混入“整数像素含两端”的加一规则。\n把两个同样大小的正方形横向错开 d，边长为 w 且 $0\\le d<w$。交集面积 $(w-d)w$，并集 $(w+d)w$，于是 IoU=$(w-d)/(w+d)$。错开一个单位时，边长 10 的框 IoU=9/11，边长 2 的框却只有 1/3。绝对坐标误差相同，几何质量不同。\n检测检查还要分清错类、漏检、框偏和重复。分数只说明模型的相对确信程度，不能直接把偏掉的框变准。若同一模型 AP50 很好而 AP75 明显下降，说明它能覆盖不少对象，但更严格位置要求暴露了问题；还应查看具体预测，才能知道误差来自偏移、大小还是同类拥挤。"},{"latex":"(w+d)w","definitionContext":"两个面积均为 4 的框，交集为 2，并集为 $4+4-2=6$，所以 IoU=1/3，而不是 1/2。分母必须是覆盖总面积，不能只用真值面积。连续坐标的宽写作 $x_2-x_1$，不要混入“整数像素含两端”的加一规则。\n把两个同样大小的正方形横向错开 d，边长为 w 且 $0\\le d<w$。交集面积 $(w-d)w$，并集 $(w+d)w$，于是 IoU=$(w-d)/(w+d)$。错开一个单位时，边长 10 的框 IoU=9/11，边长 2 的框却只有 1/3。绝对坐标误差相同，几何质量不同。\n检测检查还要分清错类、漏检、框偏和重复。分数只说明模型的相对确信程度，不能直接把偏掉的框变准。若同一模型 AP50 很好而 AP75 明显下降，说明它能覆盖不少对象，但更严格位置要求暴露了问题；还应查看具体预测，才能知道误差来自偏移、大小还是同类拥挤。"},{"latex":"(w-d)/(w+d)","definitionContext":"两个面积均为 4 的框，交集为 2，并集为 $4+4-2=6$，所以 IoU=1/3，而不是 1/2。分母必须是覆盖总面积，不能只用真值面积。连续坐标的宽写作 $x_2-x_1$，不要混入“整数像素含两端”的加一规则。\n把两个同样大小的正方形横向错开 d，边长为 w 且 $0\\le d<w$。交集面积 $(w-d)w$，并集 $(w+d)w$，于是 IoU=$(w-d)/(w+d)$。错开一个单位时，边长 10 的框 IoU=9/11，边长 2 的框却只有 1/3。绝对坐标误差相同，几何质量不同。\n检测检查还要分清错类、漏检、框偏和重复。分数只说明模型的相对确信程度，不能直接把偏掉的框变准。若同一模型 AP50 很好而 AP75 明显下降，说明它能覆盖不少对象，但更严格位置要求暴露了问题；还应查看具体预测，才能知道误差来自偏移、大小还是同类拥挤。"}],"previousBlock":"b-fa073fa3135ab9eb6527aa01","nextBlock":"b-e4876a9e17336ce7e0160b91","prerequisites":[{"id":"b-7d2547a32220bd364a2e82f6","url":"#b-7d2547a32220bd364a2e82f6","role":"section-title"},{"id":"b-23338d84ff8b6b413d733655","url":"#b-23338d84ff8b6b413d733655","role":"problem-context"}],"relatedBlocks":[{"id":"b-7d2547a32220bd364a2e82f6","url":"#b-7d2547a32220bd364a2e82f6","relation":"same-reasoning"},{"id":"b-fa073fa3135ab9eb6527aa01","url":"#b-fa073fa3135ab9eb6527aa01","relation":"same-reasoning"},{"id":"b-e4876a9e17336ce7e0160b91","url":"#b-e4876a9e17336ce7e0160b91","relation":"same-reasoning"}]},{"id":"b-e4876a9e17336ce7e0160b91","type":"p","scope":"iou-and-small-boxes","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-e4876a9e17336ce7e0160b91","content":"检测检查还要分清错类、漏检、框偏和重复。分数只说明模型的相对确信程度，不能直接把偏掉的框变准。若同一模型 AP50 很好而 AP75 明显下降，说明它能覆盖不少对象，但更严格位置要求暴露了问题；还应查看具体预测，才能知道误差来自偏移、大小还是同类拥挤。","sha256":"1ed9c49b9e5eb73a79b5965633c6d3a0f4eb6ecd61881ccb591c7de5896226b8","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":"detection-iou-scale","title":"同样偏移一个单位，为什么小框损失更大？","number":2},"sectionTitle":"同样移动一个单位，小框为什么更吃亏","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":"iou-and-small-boxes","formulas":[],"previousBlock":"b-4d87b6a7ecf1d9018df56b02","nextBlock":null,"prerequisites":[{"id":"b-7d2547a32220bd364a2e82f6","url":"#b-7d2547a32220bd364a2e82f6","role":"section-title"},{"id":"b-23338d84ff8b6b413d733655","url":"#b-23338d84ff8b6b413d733655","role":"problem-context"}],"relatedBlocks":[{"id":"b-7d2547a32220bd364a2e82f6","url":"#b-7d2547a32220bd364a2e82f6","relation":"same-reasoning"},{"id":"b-fa073fa3135ab9eb6527aa01","url":"#b-fa073fa3135ab9eb6527aa01","relation":"same-reasoning"},{"id":"b-4d87b6a7ecf1d9018df56b02","url":"#b-4d87b6a7ecf1d9018df56b02","relation":"same-reasoning"}]},{"id":"b-1881a18f1bdac2b4584025ac","type":"h2","scope":"detection-ap-ranking","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-1881a18f1bdac2b4584025ac","content":"## 沿三个预测算出 AP=5/6","sha256":"d5754d6f1cf186e66840727953370fd3b6996d8759ffb23958a4ac1dc7327237","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":"detection-ap-ranking","title":"沿三个预测算出 AP=5/6","number":3},"sectionTitle":"沿三个预测算出 AP=5/6","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-6280ccf4cb24a8b3b9baee5a","prerequisites":[{"id":"b-6280ccf4cb24a8b3b9baee5a","url":"#b-6280ccf4cb24a8b3b9baee5a","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-6280ccf4cb24a8b3b9baee5a","type":"p","scope":"detection-ap-ranking","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-6280ccf4cb24a8b3b9baee5a","content":"位置是否合格只是一步；按分数处理预测时还要为对象保留唯一匹配。","sha256":"b7a19fbbe4ac204dec861d19335bc5f0d0752a456b7fa86bd9eb3e62011f6210","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":"detection-ap-ranking","title":"沿三个预测算出 AP=5/6","number":3},"sectionTitle":"沿三个预测算出 AP=5/6","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-1881a18f1bdac2b4584025ac","nextBlock":"b-680e48950f38f34bd1867427","prerequisites":[{"id":"b-1881a18f1bdac2b4584025ac","url":"#b-1881a18f1bdac2b4584025ac","role":"section-title"}],"relatedBlocks":[]},{"id":"b-680e48950f38f34bd1867427","type":"h3","scope":"ap-ranking-example","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-680e48950f38f34bd1867427","content":"### 三个预测怎样变成 AP=5/6","sha256":"bcbcb881073dbfa272a37bec76d677c7908f0adfdf2341e99ac4edd457e37673","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":"detection-ap-ranking","title":"沿三个预测算出 AP=5/6","number":3},"sectionTitle":"三个预测怎样变成 AP=5/6","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":"ap-ranking-example","formulas":[],"previousBlock":"b-6280ccf4cb24a8b3b9baee5a","nextBlock":"b-7c5f00cfec09391321e12df1","prerequisites":[{"id":"b-1881a18f1bdac2b4584025ac","url":"#b-1881a18f1bdac2b4584025ac","role":"section-title"},{"id":"b-6280ccf4cb24a8b3b9baee5a","url":"#b-6280ccf4cb24a8b3b9baee5a","role":"problem-context"}],"relatedBlocks":[{"id":"b-7c5f00cfec09391321e12df1","url":"#b-7c5f00cfec09391321e12df1","relation":"same-reasoning"},{"id":"b-0aea64a92001f00e98b6d3f3","url":"#b-0aea64a92001f00e98b6d3f3","relation":"same-reasoning"},{"id":"b-5cf60d743e5672214001395b","url":"#b-5cf60d743e5672214001395b","relation":"same-reasoning"},{"id":"b-e7be383c891af82e100ff4b7","url":"#b-e7be383c891af82e100ff4b7","relation":"same-reasoning"}]},{"id":"b-7c5f00cfec09391321e12df1","type":"p","scope":"ap-ranking-example","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-7c5f00cfec09391321e12df1","content":"固定一个类别、一张图里两个真值对象，IoU 合格门槛为 .5。三个预测按分数递减处理，身份依次是“命中对象 A”“重复命中 A”“命中 B”。同一真值不能用两次，因此 TP/FP 序列是 TP、FP、TP。","sha256":"bdfffe03be3a24a07e121a89d5e62e7a9d6c0b5dfea1399e91bce266a771302c","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":"detection-ap-ranking","title":"沿三个预测算出 AP=5/6","number":3},"sectionTitle":"三个预测怎样变成 AP=5/6","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":"ap-ranking-example","formulas":[],"previousBlock":"b-680e48950f38f34bd1867427","nextBlock":"b-0aea64a92001f00e98b6d3f3","prerequisites":[{"id":"b-680e48950f38f34bd1867427","url":"#b-680e48950f38f34bd1867427","role":"section-title"},{"id":"b-6280ccf4cb24a8b3b9baee5a","url":"#b-6280ccf4cb24a8b3b9baee5a","role":"problem-context"}],"relatedBlocks":[{"id":"b-680e48950f38f34bd1867427","url":"#b-680e48950f38f34bd1867427","relation":"same-reasoning"},{"id":"b-0aea64a92001f00e98b6d3f3","url":"#b-0aea64a92001f00e98b6d3f3","relation":"same-reasoning"},{"id":"b-5cf60d743e5672214001395b","url":"#b-5cf60d743e5672214001395b","relation":"same-reasoning"},{"id":"b-e7be383c891af82e100ff4b7","url":"#b-e7be383c891af82e100ff4b7","relation":"same-reasoning"}]},{"id":"b-0aea64a92001f00e98b6d3f3","type":"table","scope":"ap-ranking-example","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-0aea64a92001f00e98b6d3f3","content":"| 已处理预测数 | 累计 TP | 累计 FP | precision | recall |\n| --- | --- | --- | --- | --- |\n| 1 | 1 | 0 | 1 | 1/2 |\n| 2 | 1 | 1 | 1/2 | 1/2 |\n| 3 | 2 | 1 | 2/3 | 1 |","sha256":"81aa179742271e79995d786f39ca4f5d89b2277a9542f0c4d51f371a67dd84d4","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":"detection-ap-ranking","title":"沿三个预测算出 AP=5/6","number":3},"sectionTitle":"三个预测怎样变成 AP=5/6","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":"ap-ranking-example","formulas":[],"table":{"headers":["已处理预测数","累计 TP","累计 FP","precision","recall"],"rows":[["已处理预测数","累计 TP","累计 FP","precision","recall"],["1","1","0","1","1/2"],["2","1","1","1/2","1/2"],["3","2","1","2/3","1"]]},"previousBlock":"b-7c5f00cfec09391321e12df1","nextBlock":"b-5cf60d743e5672214001395b","prerequisites":[{"id":"b-680e48950f38f34bd1867427","url":"#b-680e48950f38f34bd1867427","role":"section-title"},{"id":"b-6280ccf4cb24a8b3b9baee5a","url":"#b-6280ccf4cb24a8b3b9baee5a","role":"problem-context"}],"relatedBlocks":[{"id":"b-680e48950f38f34bd1867427","url":"#b-680e48950f38f34bd1867427","relation":"same-reasoning"},{"id":"b-7c5f00cfec09391321e12df1","url":"#b-7c5f00cfec09391321e12df1","relation":"same-reasoning"},{"id":"b-5cf60d743e5672214001395b","url":"#b-5cf60d743e5672214001395b","relation":"same-reasoning"},{"id":"b-e7be383c891af82e100ff4b7","url":"#b-e7be383c891af82e100ff4b7","relation":"same-reasoning"}]},{"id":"b-5cf60d743e5672214001395b","type":"p","scope":"ap-ranking-example","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-5cf60d743e5672214001395b","content":"采用本课程实验的 all-points precision-envelope 积分：从右向左取后续 precision 的最大值，于是包络为 1、2/3、2/3。recall 只有两次增加，各增加 1/2，所以 AP=$(1/2)\\times1+(1/2)\\times(2/3)=5/6$。把第二项重复框排到最后，先后成为 TP、TP、FP，包络积分会变为 1；这说明分数排序有实际意义。","sha256":"1abe3fad38fc2fbcdd4c804056eb02bf0d5e8a5ccc49f372f8533156f282f3d7","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":"detection-ap-ranking","title":"沿三个预测算出 AP=5/6","number":3},"sectionTitle":"三个预测怎样变成 AP=5/6","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":"ap-ranking-example","formulas":[{"latex":"(1/2)\\times1+(1/2)\\times(2/3)=5/6","definitionContext":"固定一个类别、一张图里两个真值对象，IoU 合格门槛为 .5。三个预测按分数递减处理，身份依次是“命中对象 A”“重复命中 A”“命中 B”。同一真值不能用两次，因此 TP/FP 序列是 TP、FP、TP。\n采用本课程实验的 all-points precision-envelope 积分：从右向左取后续 precision 的最大值，于是包络为 1、2/3、2/3。recall 只有两次增加，各增加 1/2，所以 AP=$(1/2)\\times1+(1/2)\\times(2/3)=5/6$。把第二项重复框排到最后，先后成为 TP、TP、FP，包络积分会变为 1；这说明分数排序有实际意义。\n这里的 AP 是指定 IoU 下的一类积分；LAB10 对两个前景类分别算，再平均，分别报告 AP50 与 AP75。COCO 还有不同 IoU、recall 采样和面积等约定，不能给这个小算例贴“COCO AP”的名称。LAB09 提供与代码一致的匹配规则，遇到一个预测接近多个真值时，要按该实现的最大 IoU 对应顺序判断。"}],"previousBlock":"b-0aea64a92001f00e98b6d3f3","nextBlock":"b-e7be383c891af82e100ff4b7","prerequisites":[{"id":"b-680e48950f38f34bd1867427","url":"#b-680e48950f38f34bd1867427","role":"section-title"},{"id":"b-6280ccf4cb24a8b3b9baee5a","url":"#b-6280ccf4cb24a8b3b9baee5a","role":"problem-context"}],"relatedBlocks":[{"id":"b-680e48950f38f34bd1867427","url":"#b-680e48950f38f34bd1867427","relation":"same-reasoning"},{"id":"b-7c5f00cfec09391321e12df1","url":"#b-7c5f00cfec09391321e12df1","relation":"same-reasoning"},{"id":"b-0aea64a92001f00e98b6d3f3","url":"#b-0aea64a92001f00e98b6d3f3","relation":"same-reasoning"},{"id":"b-e7be383c891af82e100ff4b7","url":"#b-e7be383c891af82e100ff4b7","relation":"same-reasoning"}]},{"id":"b-e7be383c891af82e100ff4b7","type":"p","scope":"ap-ranking-example","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-e7be383c891af82e100ff4b7","content":"这里的 AP 是指定 IoU 下的一类积分；LAB10 对两个前景类分别算，再平均，分别报告 AP50 与 AP75。COCO 还有不同 IoU、recall 采样和面积等约定，不能给这个小算例贴“COCO AP”的名称。LAB09 提供与代码一致的匹配规则，遇到一个预测接近多个真值时，要按该实现的最大 IoU 对应顺序判断。","sha256":"ae0bbee6493290fced1bf687bb5f343334523245c57f2285e00af558b552c39f","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":"detection-ap-ranking","title":"沿三个预测算出 AP=5/6","number":3},"sectionTitle":"三个预测怎样变成 AP=5/6","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p35"],"knowledgeBlock":"ap-ranking-example","formulas":[],"previousBlock":"b-5cf60d743e5672214001395b","nextBlock":null,"prerequisites":[{"id":"b-680e48950f38f34bd1867427","url":"#b-680e48950f38f34bd1867427","role":"section-title"},{"id":"b-6280ccf4cb24a8b3b9baee5a","url":"#b-6280ccf4cb24a8b3b9baee5a","role":"problem-context"}],"relatedBlocks":[{"id":"b-680e48950f38f34bd1867427","url":"#b-680e48950f38f34bd1867427","relation":"same-reasoning"},{"id":"b-7c5f00cfec09391321e12df1","url":"#b-7c5f00cfec09391321e12df1","relation":"same-reasoning"},{"id":"b-0aea64a92001f00e98b6d3f3","url":"#b-0aea64a92001f00e98b6d3f3","relation":"same-reasoning"},{"id":"b-5cf60d743e5672214001395b","url":"#b-5cf60d743e5672214001395b","relation":"same-reasoning"}]},{"id":"b-dbe527013e51f33f30bf87bc","type":"h2","scope":"detector-lab10-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-dbe527013e51f33f30bf87bc","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":4},"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-50efb17832df14fd94de7726","prerequisites":[{"id":"b-50efb17832df14fd94de7726","url":"#b-50efb17832df14fd94de7726","role":"problem-context"}],"relatedBlocks":[{"id":"b-f39861246e1e47fe535429ed","url":"#b-f39861246e1e47fe535429ed","relation":"figure-and-explanation"}]},{"id":"b-50efb17832df14fd94de7726","type":"p","scope":"detector-lab10-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-50efb17832df14fd94de7726","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":4},"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-dbe527013e51f33f30bf87bc","nextBlock":"b-f39861246e1e47fe535429ed","prerequisites":[{"id":"b-dbe527013e51f33f30bf87bc","url":"#b-dbe527013e51f33f30bf87bc","role":"section-title"}],"relatedBlocks":[{"id":"b-f39861246e1e47fe535429ed","url":"#b-f39861246e1e47fe535429ed","relation":"figure-and-explanation"}]},{"id":"b-f39861246e1e47fe535429ed","type":"figure","scope":"detector-lab10-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-f39861246e1e47fe535429ed","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":4},"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-50efb17832df14fd94de7726","nextBlock":"b-ba994411c61ab43105b70a14","prerequisites":[{"id":"b-dbe527013e51f33f30bf87bc","url":"#b-dbe527013e51f33f30bf87bc","role":"section-title"},{"id":"b-50efb17832df14fd94de7726","url":"#b-50efb17832df14fd94de7726","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-ba994411c61ab43105b70a14","type":"h3","scope":"detector-learning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-ba994411c61ab43105b70a14","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":4},"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-f39861246e1e47fe535429ed","nextBlock":"b-5144d7c8819a0d53f760a354","prerequisites":[{"id":"b-dbe527013e51f33f30bf87bc","url":"#b-dbe527013e51f33f30bf87bc","role":"section-title"},{"id":"b-50efb17832df14fd94de7726","url":"#b-50efb17832df14fd94de7726","role":"problem-context"}],"relatedBlocks":[{"id":"b-f39861246e1e47fe535429ed","url":"#b-f39861246e1e47fe535429ed","relation":"figure-and-explanation"},{"id":"b-5144d7c8819a0d53f760a354","url":"#b-5144d7c8819a0d53f760a354","relation":"same-reasoning"},{"id":"b-eef4efc56853161c731ef41c","url":"#b-eef4efc56853161c731ef41c","relation":"same-reasoning"},{"id":"b-909d02131980053ebaa30790","url":"#b-909d02131980053ebaa30790","relation":"same-reasoning"},{"id":"b-8797fcf5aed1bdbae216334f","url":"#b-8797fcf5aed1bdbae216334f","relation":"same-reasoning"},{"id":"b-8a8034b89aed132313d436b1","url":"#b-8a8034b89aed132313d436b1","relation":"same-reasoning"}]},{"id":"b-5144d7c8819a0d53f760a354","type":"p","scope":"detector-learning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-5144d7c8819a0d53f760a354","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":4},"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-ba994411c61ab43105b70a14","nextBlock":"b-eef4efc56853161c731ef41c","prerequisites":[{"id":"b-ba994411c61ab43105b70a14","url":"#b-ba994411c61ab43105b70a14","role":"section-title"},{"id":"b-50efb17832df14fd94de7726","url":"#b-50efb17832df14fd94de7726","role":"problem-context"}],"relatedBlocks":[{"id":"b-f39861246e1e47fe535429ed","url":"#b-f39861246e1e47fe535429ed","relation":"figure-and-explanation"},{"id":"b-ba994411c61ab43105b70a14","url":"#b-ba994411c61ab43105b70a14","relation":"same-reasoning"},{"id":"b-eef4efc56853161c731ef41c","url":"#b-eef4efc56853161c731ef41c","relation":"same-reasoning"},{"id":"b-909d02131980053ebaa30790","url":"#b-909d02131980053ebaa30790","relation":"same-reasoning"},{"id":"b-8797fcf5aed1bdbae216334f","url":"#b-8797fcf5aed1bdbae216334f","relation":"same-reasoning"},{"id":"b-8a8034b89aed132313d436b1","url":"#b-8a8034b89aed132313d436b1","relation":"same-reasoning"}]},{"id":"b-eef4efc56853161c731ef41c","type":"p","scope":"detector-learning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-eef4efc56853161c731ef41c","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":4},"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-5144d7c8819a0d53f760a354","nextBlock":"b-909d02131980053ebaa30790","prerequisites":[{"id":"b-ba994411c61ab43105b70a14","url":"#b-ba994411c61ab43105b70a14","role":"section-title"},{"id":"b-50efb17832df14fd94de7726","url":"#b-50efb17832df14fd94de7726","role":"problem-context"}],"relatedBlocks":[{"id":"b-f39861246e1e47fe535429ed","url":"#b-f39861246e1e47fe535429ed","relation":"figure-and-explanation"},{"id":"b-ba994411c61ab43105b70a14","url":"#b-ba994411c61ab43105b70a14","relation":"same-reasoning"},{"id":"b-5144d7c8819a0d53f760a354","url":"#b-5144d7c8819a0d53f760a354","relation":"same-reasoning"},{"id":"b-909d02131980053ebaa30790","url":"#b-909d02131980053ebaa30790","relation":"same-reasoning"},{"id":"b-8797fcf5aed1bdbae216334f","url":"#b-8797fcf5aed1bdbae216334f","relation":"same-reasoning"},{"id":"b-8a8034b89aed132313d436b1","url":"#b-8a8034b89aed132313d436b1","relation":"same-reasoning"}]},{"id":"b-909d02131980053ebaa30790","type":"table","scope":"detector-learning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-909d02131980053ebaa30790","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":4},"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-eef4efc56853161c731ef41c","nextBlock":"b-8797fcf5aed1bdbae216334f","prerequisites":[{"id":"b-ba994411c61ab43105b70a14","url":"#b-ba994411c61ab43105b70a14","role":"section-title"},{"id":"b-50efb17832df14fd94de7726","url":"#b-50efb17832df14fd94de7726","role":"problem-context"}],"relatedBlocks":[{"id":"b-f39861246e1e47fe535429ed","url":"#b-f39861246e1e47fe535429ed","relation":"figure-and-explanation"},{"id":"b-ba994411c61ab43105b70a14","url":"#b-ba994411c61ab43105b70a14","relation":"same-reasoning"},{"id":"b-5144d7c8819a0d53f760a354","url":"#b-5144d7c8819a0d53f760a354","relation":"same-reasoning"},{"id":"b-eef4efc56853161c731ef41c","url":"#b-eef4efc56853161c731ef41c","relation":"same-reasoning"},{"id":"b-8797fcf5aed1bdbae216334f","url":"#b-8797fcf5aed1bdbae216334f","relation":"same-reasoning"},{"id":"b-8a8034b89aed132313d436b1","url":"#b-8a8034b89aed132313d436b1","relation":"same-reasoning"}]},{"id":"b-8797fcf5aed1bdbae216334f","type":"p","scope":"detector-learning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-8797fcf5aed1bdbae216334f","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":4},"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-909d02131980053ebaa30790","nextBlock":"b-8a8034b89aed132313d436b1","prerequisites":[{"id":"b-ba994411c61ab43105b70a14","url":"#b-ba994411c61ab43105b70a14","role":"section-title"},{"id":"b-50efb17832df14fd94de7726","url":"#b-50efb17832df14fd94de7726","role":"problem-context"}],"relatedBlocks":[{"id":"b-f39861246e1e47fe535429ed","url":"#b-f39861246e1e47fe535429ed","relation":"figure-and-explanation"},{"id":"b-ba994411c61ab43105b70a14","url":"#b-ba994411c61ab43105b70a14","relation":"same-reasoning"},{"id":"b-5144d7c8819a0d53f760a354","url":"#b-5144d7c8819a0d53f760a354","relation":"same-reasoning"},{"id":"b-eef4efc56853161c731ef41c","url":"#b-eef4efc56853161c731ef41c","relation":"same-reasoning"},{"id":"b-909d02131980053ebaa30790","url":"#b-909d02131980053ebaa30790","relation":"same-reasoning"},{"id":"b-8a8034b89aed132313d436b1","url":"#b-8a8034b89aed132313d436b1","relation":"same-reasoning"}]},{"id":"b-8a8034b89aed132313d436b1","type":"p","scope":"detector-learning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-8a8034b89aed132313d436b1","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":4},"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-8797fcf5aed1bdbae216334f","nextBlock":null,"prerequisites":[{"id":"b-ba994411c61ab43105b70a14","url":"#b-ba994411c61ab43105b70a14","role":"section-title"},{"id":"b-50efb17832df14fd94de7726","url":"#b-50efb17832df14fd94de7726","role":"problem-context"}],"relatedBlocks":[{"id":"b-f39861246e1e47fe535429ed","url":"#b-f39861246e1e47fe535429ed","relation":"figure-and-explanation"},{"id":"b-ba994411c61ab43105b70a14","url":"#b-ba994411c61ab43105b70a14","relation":"same-reasoning"},{"id":"b-5144d7c8819a0d53f760a354","url":"#b-5144d7c8819a0d53f760a354","relation":"same-reasoning"},{"id":"b-eef4efc56853161c731ef41c","url":"#b-eef4efc56853161c731ef41c","relation":"same-reasoning"},{"id":"b-909d02131980053ebaa30790","url":"#b-909d02131980053ebaa30790","relation":"same-reasoning"},{"id":"b-8797fcf5aed1bdbae216334f","url":"#b-8797fcf5aed1bdbae216334f","relation":"same-reasoning"}]},{"id":"b-752f624db54b5c1ba4df738e","type":"h2","scope":"detector-nms-measured-tradeoff","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-752f624db54b5c1ba4df738e","content":"## 只改变 NMS，检查新增框究竟是什么","sha256":"ad486ca4482b94aae52a9ea4f8d905fe7e5e24dd0d4d0db64e99b64896a13528","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-nms-measured-tradeoff","title":"只改变 NMS，检查新增框究竟是什么","number":5},"sectionTitle":"只改变 NMS，检查新增框究竟是什么","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p34","l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-734e68dc13afaf39bbba5d88","prerequisites":[{"id":"b-734e68dc13afaf39bbba5d88","url":"#b-734e68dc13afaf39bbba5d88","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-734e68dc13afaf39bbba5d88","type":"p","scope":"detector-nms-measured-tradeoff","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-734e68dc13afaf39bbba5d88","content":"已有模型让我们能够固定输入与权重，只检查后处理带来的取舍。","sha256":"ae3b525af2a0705eadc8d05ceedf8fd75aeadbabd67ee7903b401f0b798bdcea","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-nms-measured-tradeoff","title":"只改变 NMS，检查新增框究竟是什么","number":5},"sectionTitle":"只改变 NMS，检查新增框究竟是什么","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p34","l3-r2-p35"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-752f624db54b5c1ba4df738e","nextBlock":"b-446addbdf40a6eb3aeae1349","prerequisites":[{"id":"b-752f624db54b5c1ba4df738e","url":"#b-752f624db54b5c1ba4df738e","role":"section-title"}],"relatedBlocks":[]},{"id":"b-446addbdf40a6eb3aeae1349","type":"h3","scope":"nms-measured-tradeoff","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-446addbdf40a6eb3aeae1349","content":"### 同一检测器改变 NMS，召回为何变好而精度变差","sha256":"3264b99394a49764a79ebd0c08e1962708be37e7d92211f58fd356b70899c02f","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-nms-measured-tradeoff","title":"只改变 NMS，检查新增框究竟是什么","number":5},"sectionTitle":"同一检测器改变 NMS，召回为何变好而精度变差","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p34","l3-r2-p35"],"knowledgeBlock":"nms-measured-tradeoff","formulas":[],"previousBlock":"b-734e68dc13afaf39bbba5d88","nextBlock":"b-66e196112a2acbfcaf290c80","prerequisites":[{"id":"b-752f624db54b5c1ba4df738e","url":"#b-752f624db54b5c1ba4df738e","role":"section-title"},{"id":"b-734e68dc13afaf39bbba5d88","url":"#b-734e68dc13afaf39bbba5d88","role":"problem-context"}],"relatedBlocks":[{"id":"b-66e196112a2acbfcaf290c80","url":"#b-66e196112a2acbfcaf290c80","relation":"same-reasoning"},{"id":"b-ccca693e4d6e131b7f715424","url":"#b-ccca693e4d6e131b7f715424","relation":"same-reasoning"},{"id":"b-afcfdf9f9a37423b01971fcc","url":"#b-afcfdf9f9a37423b01971fcc","relation":"same-reasoning"},{"id":"b-ca9d527c53cd6cd10ea763d4","url":"#b-ca9d527c53cd6cd10ea763d4","relation":"same-reasoning"}]},{"id":"b-66e196112a2acbfcaf290c80","type":"p","scope":"nms-measured-tradeoff","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-66e196112a2acbfcaf290c80","content":"LAB10 已保存一组更接近真实使用的对照：输入与训练后权重保持不变，只把 NMS IoU 阈值从 .35 放宽到 .65。展示/工作点的分数门槛固定 .4。","sha256":"8b0e46c250ed7908b144b4e40151bb579c49ca95fce5168dc35905e6cd61fbbb","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-nms-measured-tradeoff","title":"只改变 NMS，检查新增框究竟是什么","number":5},"sectionTitle":"同一检测器改变 NMS，召回为何变好而精度变差","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p34","l3-r2-p35"],"knowledgeBlock":"nms-measured-tradeoff","formulas":[],"previousBlock":"b-446addbdf40a6eb3aeae1349","nextBlock":"b-ccca693e4d6e131b7f715424","prerequisites":[{"id":"b-446addbdf40a6eb3aeae1349","url":"#b-446addbdf40a6eb3aeae1349","role":"section-title"},{"id":"b-734e68dc13afaf39bbba5d88","url":"#b-734e68dc13afaf39bbba5d88","role":"problem-context"}],"relatedBlocks":[{"id":"b-446addbdf40a6eb3aeae1349","url":"#b-446addbdf40a6eb3aeae1349","relation":"same-reasoning"},{"id":"b-ccca693e4d6e131b7f715424","url":"#b-ccca693e4d6e131b7f715424","relation":"same-reasoning"},{"id":"b-afcfdf9f9a37423b01971fcc","url":"#b-afcfdf9f9a37423b01971fcc","relation":"same-reasoning"},{"id":"b-ca9d527c53cd6cd10ea763d4","url":"#b-ca9d527c53cd6cd10ea763d4","relation":"same-reasoning"}]},{"id":"b-ccca693e4d6e131b7f715424","type":"table","scope":"nms-measured-tradeoff","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-ccca693e4d6e131b7f715424","content":"| NMS IoU | recall@工作点 | precision@工作点 | AP50 |\n| --- | --- | --- | --- |\n| .35 | .98311 | .94175 | .97303 |\n| .65 | 1.00000 | .41926 | .95866 |","sha256":"6568df1908194db0667c411a0d52480b9be36f1ec5c80e3aa45bba20e564fa18","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-nms-measured-tradeoff","title":"只改变 NMS，检查新增框究竟是什么","number":5},"sectionTitle":"同一检测器改变 NMS，召回为何变好而精度变差","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p34","l3-r2-p35"],"knowledgeBlock":"nms-measured-tradeoff","formulas":[],"table":{"headers":["NMS IoU","recall@工作点","precision@工作点","AP50"],"rows":[["NMS IoU","recall@工作点","precision@工作点","AP50"],[".35",".98311",".94175",".97303"],[".65","1.00000",".41926",".95866"]]},"previousBlock":"b-66e196112a2acbfcaf290c80","nextBlock":"b-afcfdf9f9a37423b01971fcc","prerequisites":[{"id":"b-446addbdf40a6eb3aeae1349","url":"#b-446addbdf40a6eb3aeae1349","role":"section-title"},{"id":"b-734e68dc13afaf39bbba5d88","url":"#b-734e68dc13afaf39bbba5d88","role":"problem-context"}],"relatedBlocks":[{"id":"b-446addbdf40a6eb3aeae1349","url":"#b-446addbdf40a6eb3aeae1349","relation":"same-reasoning"},{"id":"b-66e196112a2acbfcaf290c80","url":"#b-66e196112a2acbfcaf290c80","relation":"same-reasoning"},{"id":"b-afcfdf9f9a37423b01971fcc","url":"#b-afcfdf9f9a37423b01971fcc","relation":"same-reasoning"},{"id":"b-ca9d527c53cd6cd10ea763d4","url":"#b-ca9d527c53cd6cd10ea763d4","relation":"same-reasoning"}]},{"id":"b-afcfdf9f9a37423b01971fcc","type":"p","scope":"nms-measured-tradeoff","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-afcfdf9f9a37423b01971fcc","content":"放宽后，一些原来被抑制的候选重新出现，补到了最后少数对象；同时大量同一对象的重复框留在输出，所以 precision 大降。AP 的变化比这个固定工作点的 precision 小，因为 AP 汇总整个分数排序，而工作点只看分数至少 .4 的输出。","sha256":"fa71bd1d1c49b9fd714195acb6b40c28ccf5d6455b9dd9408b7c6e7101b36997","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-nms-measured-tradeoff","title":"只改变 NMS，检查新增框究竟是什么","number":5},"sectionTitle":"同一检测器改变 NMS，召回为何变好而精度变差","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p34","l3-r2-p35"],"knowledgeBlock":"nms-measured-tradeoff","formulas":[],"previousBlock":"b-ccca693e4d6e131b7f715424","nextBlock":"b-ca9d527c53cd6cd10ea763d4","prerequisites":[{"id":"b-446addbdf40a6eb3aeae1349","url":"#b-446addbdf40a6eb3aeae1349","role":"section-title"},{"id":"b-734e68dc13afaf39bbba5d88","url":"#b-734e68dc13afaf39bbba5d88","role":"problem-context"}],"relatedBlocks":[{"id":"b-446addbdf40a6eb3aeae1349","url":"#b-446addbdf40a6eb3aeae1349","relation":"same-reasoning"},{"id":"b-66e196112a2acbfcaf290c80","url":"#b-66e196112a2acbfcaf290c80","relation":"same-reasoning"},{"id":"b-ccca693e4d6e131b7f715424","url":"#b-ccca693e4d6e131b7f715424","relation":"same-reasoning"},{"id":"b-ca9d527c53cd6cd10ea763d4","url":"#b-ca9d527c53cd6cd10ea763d4","relation":"same-reasoning"}]},{"id":"b-ca9d527c53cd6cd10ea763d4","type":"p","scope":"nms-measured-tradeoff","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-ca9d527c53cd6cd10ea763d4","content":"这组已有本地结果不是本轮新运行。它的价值是把 P34 的两种效应落实到同一模型：预测数量增加本身没有确定好坏，必须看新增的是哪个对象、是否重复，以及部署时怎样使用阈值。选阈值应使用验证集与任务成本；不能反复挑保留集最漂亮的数字再称泛化结果。","sha256":"07c3598d8f94057887eadde2d224c44388ad519e293cfc2ead957ce4985de933","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-nms-measured-tradeoff","title":"只改变 NMS，检查新增框究竟是什么","number":5},"sectionTitle":"同一检测器改变 NMS，召回为何变好而精度变差","currentQuestion":"怎样从错误图走到 IoU、排序匹配与已有检测实验？","sourceSlides":["l3-r2-p34","l3-r2-p35"],"knowledgeBlock":"nms-measured-tradeoff","formulas":[],"previousBlock":"b-afcfdf9f9a37423b01971fcc","nextBlock":null,"prerequisites":[{"id":"b-446addbdf40a6eb3aeae1349","url":"#b-446addbdf40a6eb3aeae1349","role":"section-title"},{"id":"b-734e68dc13afaf39bbba5d88","url":"#b-734e68dc13afaf39bbba5d88","role":"problem-context"}],"relatedBlocks":[{"id":"b-446addbdf40a6eb3aeae1349","url":"#b-446addbdf40a6eb3aeae1349","relation":"same-reasoning"},{"id":"b-66e196112a2acbfcaf290c80","url":"#b-66e196112a2acbfcaf290c80","relation":"same-reasoning"},{"id":"b-ccca693e4d6e131b7f715424","url":"#b-ccca693e4d6e131b7f715424","relation":"same-reasoning"},{"id":"b-afcfdf9f9a37423b01971fcc","url":"#b-afcfdf9f9a37423b01971fcc","relation":"same-reasoning"}]},{"id":"b-2bded4a5236159ca45bd550b","type":"h2","scope":"dense-detector-shared-problems","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-2bded4a5236159ca45bd550b","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":6},"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-80d1985c1080b8f662880e27","prerequisites":[{"id":"b-80d1985c1080b8f662880e27","url":"#b-80d1985c1080b8f662880e27","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-80d1985c1080b8f662880e27","type":"p","scope":"dense-detector-shared-problems","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-80d1985c1080b8f662880e27","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":6},"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-2bded4a5236159ca45bd550b","nextBlock":"b-05bd3a1b982fcbc8e92accc2","prerequisites":[{"id":"b-2bded4a5236159ca45bd550b","url":"#b-2bded4a5236159ca45bd550b","role":"section-title"}],"relatedBlocks":[]},{"id":"b-05bd3a1b982fcbc8e92accc2","type":"h3","scope":"dense-detector-ssd-extension","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-05bd3a1b982fcbc8e92accc2","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":6},"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-80d1985c1080b8f662880e27","nextBlock":"b-2cb214361a28884d64d084cc","prerequisites":[{"id":"b-2bded4a5236159ca45bd550b","url":"#b-2bded4a5236159ca45bd550b","role":"section-title"},{"id":"b-80d1985c1080b8f662880e27","url":"#b-80d1985c1080b8f662880e27","role":"problem-context"}],"relatedBlocks":[{"id":"b-2cb214361a28884d64d084cc","url":"#b-2cb214361a28884d64d084cc","relation":"same-reasoning"},{"id":"b-c44aec173f967a11dbeb0c30","url":"#b-c44aec173f967a11dbeb0c30","relation":"same-reasoning"},{"id":"b-987676769e84a9fcbdaabe6f","url":"#b-987676769e84a9fcbdaabe6f","relation":"same-reasoning"},{"id":"b-e7412dd60f7e72eb6d7a2b81","url":"#b-e7412dd60f7e72eb6d7a2b81","relation":"same-reasoning"}]},{"id":"b-2cb214361a28884d64d084cc","type":"p","scope":"dense-detector-ssd-extension","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-2cb214361a28884d64d084cc","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":6},"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-05bd3a1b982fcbc8e92accc2","nextBlock":"b-c44aec173f967a11dbeb0c30","prerequisites":[{"id":"b-05bd3a1b982fcbc8e92accc2","url":"#b-05bd3a1b982fcbc8e92accc2","role":"section-title"},{"id":"b-80d1985c1080b8f662880e27","url":"#b-80d1985c1080b8f662880e27","role":"problem-context"}],"relatedBlocks":[{"id":"b-05bd3a1b982fcbc8e92accc2","url":"#b-05bd3a1b982fcbc8e92accc2","relation":"same-reasoning"},{"id":"b-c44aec173f967a11dbeb0c30","url":"#b-c44aec173f967a11dbeb0c30","relation":"same-reasoning"},{"id":"b-987676769e84a9fcbdaabe6f","url":"#b-987676769e84a9fcbdaabe6f","relation":"same-reasoning"},{"id":"b-e7412dd60f7e72eb6d7a2b81","url":"#b-e7412dd60f7e72eb6d7a2b81","relation":"same-reasoning"}]},{"id":"b-c44aec173f967a11dbeb0c30","type":"p","scope":"dense-detector-ssd-extension","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-c44aec173f967a11dbeb0c30","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":6},"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-2cb214361a28884d64d084cc","nextBlock":"b-987676769e84a9fcbdaabe6f","prerequisites":[{"id":"b-05bd3a1b982fcbc8e92accc2","url":"#b-05bd3a1b982fcbc8e92accc2","role":"section-title"},{"id":"b-80d1985c1080b8f662880e27","url":"#b-80d1985c1080b8f662880e27","role":"problem-context"}],"relatedBlocks":[{"id":"b-05bd3a1b982fcbc8e92accc2","url":"#b-05bd3a1b982fcbc8e92accc2","relation":"same-reasoning"},{"id":"b-2cb214361a28884d64d084cc","url":"#b-2cb214361a28884d64d084cc","relation":"same-reasoning"},{"id":"b-987676769e84a9fcbdaabe6f","url":"#b-987676769e84a9fcbdaabe6f","relation":"same-reasoning"},{"id":"b-e7412dd60f7e72eb6d7a2b81","url":"#b-e7412dd60f7e72eb6d7a2b81","relation":"same-reasoning"}]},{"id":"b-987676769e84a9fcbdaabe6f","type":"p","scope":"dense-detector-ssd-extension","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-987676769e84a9fcbdaabe6f","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":6},"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-c44aec173f967a11dbeb0c30","nextBlock":"b-e7412dd60f7e72eb6d7a2b81","prerequisites":[{"id":"b-05bd3a1b982fcbc8e92accc2","url":"#b-05bd3a1b982fcbc8e92accc2","role":"section-title"},{"id":"b-80d1985c1080b8f662880e27","url":"#b-80d1985c1080b8f662880e27","role":"problem-context"}],"relatedBlocks":[{"id":"b-05bd3a1b982fcbc8e92accc2","url":"#b-05bd3a1b982fcbc8e92accc2","relation":"same-reasoning"},{"id":"b-2cb214361a28884d64d084cc","url":"#b-2cb214361a28884d64d084cc","relation":"same-reasoning"},{"id":"b-c44aec173f967a11dbeb0c30","url":"#b-c44aec173f967a11dbeb0c30","relation":"same-reasoning"},{"id":"b-e7412dd60f7e72eb6d7a2b81","url":"#b-e7412dd60f7e72eb6d7a2b81","relation":"same-reasoning"}]},{"id":"b-e7412dd60f7e72eb6d7a2b81","type":"p","scope":"dense-detector-ssd-extension","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-e7412dd60f7e72eb6d7a2b81","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":6},"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-987676769e84a9fcbdaabe6f","nextBlock":null,"prerequisites":[{"id":"b-05bd3a1b982fcbc8e92accc2","url":"#b-05bd3a1b982fcbc8e92accc2","role":"section-title"},{"id":"b-80d1985c1080b8f662880e27","url":"#b-80d1985c1080b8f662880e27","role":"problem-context"}],"relatedBlocks":[{"id":"b-05bd3a1b982fcbc8e92accc2","url":"#b-05bd3a1b982fcbc8e92accc2","relation":"same-reasoning"},{"id":"b-2cb214361a28884d64d084cc","url":"#b-2cb214361a28884d64d084cc","relation":"same-reasoning"},{"id":"b-c44aec173f967a11dbeb0c30","url":"#b-c44aec173f967a11dbeb0c30","relation":"same-reasoning"},{"id":"b-987676769e84a9fcbdaabe6f","url":"#b-987676769e84a9fcbdaabe6f","relation":"same-reasoning"}]},{"id":"b-bc9124cedc5b09deedc151cd","type":"h2","scope":"focal-weighting-and-assignment","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-bc9124cedc5b09deedc151cd","content":"## 把难例权重与对象匹配分成两件事","sha256":"aa1b2d370c18fda3ba4bcd814dc219852d6bfae0d1e55d9566c20d15cb41a81b","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":"focal-weighting-and-assignment","title":"把难例权重与对象匹配分成两件事","number":7},"sectionTitle":"把难例权重与对象匹配分成两件事","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-p35","l3-r2-p38"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-079b27913a200169a4693cf2","prerequisites":[{"id":"b-079b27913a200169a4693cf2","url":"#b-079b27913a200169a4693cf2","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-079b27913a200169a4693cf2","type":"p","scope":"focal-weighting-and-assignment","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-079b27913a200169a4693cf2","content":"一种训练调整影响样本权重，另一种先确定预测与目标的对应。","sha256":"5f80dffbeba94a984e51bdada1fcbd4fb37ede055ba58977630c058c6cb4aa5c","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":"focal-weighting-and-assignment","title":"把难例权重与对象匹配分成两件事","number":7},"sectionTitle":"把难例权重与对象匹配分成两件事","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-p35","l3-r2-p38"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-bc9124cedc5b09deedc151cd","nextBlock":"b-8bde3aebadeb8dbb78bfaa57","prerequisites":[{"id":"b-bc9124cedc5b09deedc151cd","url":"#b-bc9124cedc5b09deedc151cd","role":"section-title"}],"relatedBlocks":[]},{"id":"b-8bde3aebadeb8dbb78bfaa57","type":"h3","scope":"focal-versus-assignment","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-8bde3aebadeb8dbb78bfaa57","content":"### 选读：难例权重和对象匹配不是一件事","sha256":"2ae031310e8b7ace430d6c798354cdf0078d2c2f282436227dab57163f170b1b","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":"focal-weighting-and-assignment","title":"把难例权重与对象匹配分成两件事","number":7},"sectionTitle":"选读：难例权重和对象匹配不是一件事","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-p35","l3-r2-p38"],"knowledgeBlock":"focal-versus-assignment","formulas":[],"previousBlock":"b-079b27913a200169a4693cf2","nextBlock":"b-b0ea4c6e23604dba4fc4ba7e","prerequisites":[{"id":"b-bc9124cedc5b09deedc151cd","url":"#b-bc9124cedc5b09deedc151cd","role":"section-title"},{"id":"b-079b27913a200169a4693cf2","url":"#b-079b27913a200169a4693cf2","role":"problem-context"}],"relatedBlocks":[{"id":"b-b0ea4c6e23604dba4fc4ba7e","url":"#b-b0ea4c6e23604dba4fc4ba7e","relation":"same-reasoning"},{"id":"b-4120daf249592055a1d25758","url":"#b-4120daf249592055a1d25758","relation":"same-reasoning"},{"id":"b-263e596b99b06ddd6280e3a2","url":"#b-263e596b99b06ddd6280e3a2","relation":"same-reasoning"}]},{"id":"b-b0ea4c6e23604dba4fc4ba7e","type":"p","scope":"focal-versus-assignment","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-b0ea4c6e23604dba4fc4ba7e","content":"大量背景参考框已经很容易分类，却可能因数量占优而贡献大量总损失。focal loss 在基本交叉熵外乘 $(1-p_t)^\\gamma$。取 γ=2，正确类概率 .9 的样本权重 .01，概率 .5 的权重 .25，后者相对贡献更大。这是在调整哪些样本值得继续关注，不直接增加空间信息，也不指定哪项预测负责哪只对象。","sha256":"acad29e92db22a1ad00a309903470a4328f8c18cfdf4fac20c3446621f0f6335","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":"focal-weighting-and-assignment","title":"把难例权重与对象匹配分成两件事","number":7},"sectionTitle":"选读：难例权重和对象匹配不是一件事","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-p35","l3-r2-p38"],"knowledgeBlock":"focal-versus-assignment","formulas":[{"latex":"(1-p_t)^\\gamma","definitionContext":"大量背景参考框已经很容易分类，却可能因数量占优而贡献大量总损失。focal loss 在基本交叉熵外乘 $(1-p_t)^\\gamma$。取 γ=2，正确类概率 .9 的样本权重 .01，概率 .5 的权重 .25，后者相对贡献更大。这是在调整哪些样本值得继续关注，不直接增加空间信息，也不指定哪项预测负责哪只对象。\n对象匹配回答另一件事：当真值是一组无序对象，预测槽位该怎样对应它们？给两预测、两真值的代价矩阵 `[[2,.3],[.4,2]]`，对角总代价 4，交叉 .7，所以选择交叉。交换标注数组顺序不该改变对象集合本身，匹配用于消除这个任意顺序。\nLAB20 从类别概率、框 L1 与 IoU 实际构造成本，再穷举 2! 分配；它是足够透明的解析练习。这里没有图像编码器学习，也不是完整 DETR。主讲保留对象身份的必要性，这些损失组织细节进入选读即可，不必为保持课程丰富而重新开一章。"}],"previousBlock":"b-8bde3aebadeb8dbb78bfaa57","nextBlock":"b-4120daf249592055a1d25758","prerequisites":[{"id":"b-8bde3aebadeb8dbb78bfaa57","url":"#b-8bde3aebadeb8dbb78bfaa57","role":"section-title"},{"id":"b-079b27913a200169a4693cf2","url":"#b-079b27913a200169a4693cf2","role":"problem-context"}],"relatedBlocks":[{"id":"b-8bde3aebadeb8dbb78bfaa57","url":"#b-8bde3aebadeb8dbb78bfaa57","relation":"same-reasoning"},{"id":"b-4120daf249592055a1d25758","url":"#b-4120daf249592055a1d25758","relation":"same-reasoning"},{"id":"b-263e596b99b06ddd6280e3a2","url":"#b-263e596b99b06ddd6280e3a2","relation":"same-reasoning"}]},{"id":"b-4120daf249592055a1d25758","type":"p","scope":"focal-versus-assignment","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-4120daf249592055a1d25758","content":"对象匹配回答另一件事：当真值是一组无序对象，预测槽位该怎样对应它们？给两预测、两真值的代价矩阵 `[[2,.3],[.4,2]]`，对角总代价 4，交叉 .7，所以选择交叉。交换标注数组顺序不该改变对象集合本身，匹配用于消除这个任意顺序。","sha256":"6d41a42fe14ef8b81dbbf4c540876335e4e04e4ecf75f45c1c416ef808b7f693","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":"focal-weighting-and-assignment","title":"把难例权重与对象匹配分成两件事","number":7},"sectionTitle":"选读：难例权重和对象匹配不是一件事","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-p35","l3-r2-p38"],"knowledgeBlock":"focal-versus-assignment","formulas":[],"previousBlock":"b-b0ea4c6e23604dba4fc4ba7e","nextBlock":"b-263e596b99b06ddd6280e3a2","prerequisites":[{"id":"b-8bde3aebadeb8dbb78bfaa57","url":"#b-8bde3aebadeb8dbb78bfaa57","role":"section-title"},{"id":"b-079b27913a200169a4693cf2","url":"#b-079b27913a200169a4693cf2","role":"problem-context"}],"relatedBlocks":[{"id":"b-8bde3aebadeb8dbb78bfaa57","url":"#b-8bde3aebadeb8dbb78bfaa57","relation":"same-reasoning"},{"id":"b-b0ea4c6e23604dba4fc4ba7e","url":"#b-b0ea4c6e23604dba4fc4ba7e","relation":"same-reasoning"},{"id":"b-263e596b99b06ddd6280e3a2","url":"#b-263e596b99b06ddd6280e3a2","relation":"same-reasoning"}]},{"id":"b-263e596b99b06ddd6280e3a2","type":"p","scope":"focal-versus-assignment","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-263e596b99b06ddd6280e3a2","content":"LAB20 从类别概率、框 L1 与 IoU 实际构造成本，再穷举 2! 分配；它是足够透明的解析练习。这里没有图像编码器学习，也不是完整 DETR。主讲保留对象身份的必要性，这些损失组织细节进入选读即可，不必为保持课程丰富而重新开一章。","sha256":"09c26eb8dd7894d9423b7b2465780fda6811ebe147e75d5d917afc5f99951c94","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":"focal-weighting-and-assignment","title":"把难例权重与对象匹配分成两件事","number":7},"sectionTitle":"选读：难例权重和对象匹配不是一件事","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-p35","l3-r2-p38"],"knowledgeBlock":"focal-versus-assignment","formulas":[],"previousBlock":"b-4120daf249592055a1d25758","nextBlock":null,"prerequisites":[{"id":"b-8bde3aebadeb8dbb78bfaa57","url":"#b-8bde3aebadeb8dbb78bfaa57","role":"section-title"},{"id":"b-079b27913a200169a4693cf2","url":"#b-079b27913a200169a4693cf2","role":"problem-context"}],"relatedBlocks":[{"id":"b-8bde3aebadeb8dbb78bfaa57","url":"#b-8bde3aebadeb8dbb78bfaa57","relation":"same-reasoning"},{"id":"b-b0ea4c6e23604dba4fc4ba7e","url":"#b-b0ea4c6e23604dba4fc4ba7e","relation":"same-reasoning"},{"id":"b-4120daf249592055a1d25758","url":"#b-4120daf249592055a1d25758","relation":"same-reasoning"}]},{"id":"b-b7e45148839ce1d171344d73","type":"h2","scope":"set-matching-lab20-costs","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-b7e45148839ce1d171344d73","content":"## 选读 LAB20：沿真实输入核对交叉匹配","sha256":"4726920945ac0b498f7801919906ebe166fc4f0bd9699b905e0a52247f1168e2","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":"set-matching-lab20-costs","title":"选读 LAB20：沿真实输入核对交叉匹配","number":8},"sectionTitle":"选读 LAB20：沿真实输入核对交叉匹配","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-p35","l3-r2-p39"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-e637ab72577ef03dfac4f03d","prerequisites":[{"id":"b-e637ab72577ef03dfac4f03d","url":"#b-e637ab72577ef03dfac4f03d","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-e637ab72577ef03dfac4f03d","type":"p","scope":"set-matching-lab20-costs","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-e637ab72577ef03dfac4f03d","content":"用已保存的小算例落实上一页的对应问题，检查成本与数组顺序。","sha256":"80aed3648cc101c5871ace40ae8fba4028b3be912a2e368644520e77b31c275c","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":"set-matching-lab20-costs","title":"选读 LAB20：沿真实输入核对交叉匹配","number":8},"sectionTitle":"选读 LAB20：沿真实输入核对交叉匹配","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-p35","l3-r2-p39"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b7e45148839ce1d171344d73","nextBlock":"b-07496f85b77ac1623307e2eb","prerequisites":[{"id":"b-b7e45148839ce1d171344d73","url":"#b-b7e45148839ce1d171344d73","role":"section-title"}],"relatedBlocks":[]},{"id":"b-07496f85b77ac1623307e2eb","type":"h3","scope":"set-matching-real-inputs","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-07496f85b77ac1623307e2eb","content":"### LAB20 的两个槽位为何应交叉匹配","sha256":"e22ba271dbf8de903d354b548af45e31549397c87cf88aa2ff55aab69ec47377","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":"set-matching-lab20-costs","title":"选读 LAB20：沿真实输入核对交叉匹配","number":8},"sectionTitle":"LAB20 的两个槽位为何应交叉匹配","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-p35","l3-r2-p39"],"knowledgeBlock":"set-matching-real-inputs","formulas":[],"previousBlock":"b-e637ab72577ef03dfac4f03d","nextBlock":"b-0f92d758c53ae5d4a6f20de3","prerequisites":[{"id":"b-b7e45148839ce1d171344d73","url":"#b-b7e45148839ce1d171344d73","role":"section-title"},{"id":"b-e637ab72577ef03dfac4f03d","url":"#b-e637ab72577ef03dfac4f03d","role":"problem-context"}],"relatedBlocks":[{"id":"b-0f92d758c53ae5d4a6f20de3","url":"#b-0f92d758c53ae5d4a6f20de3","relation":"same-reasoning"},{"id":"b-91a5f7f6a09aaa85af8addb8","url":"#b-91a5f7f6a09aaa85af8addb8","relation":"same-reasoning"},{"id":"b-efb5f1232e4b79525f38a178","url":"#b-efb5f1232e4b79525f38a178","relation":"same-reasoning"},{"id":"b-2d6fed637b1b9ac05025df8c","url":"#b-2d6fed637b1b9ac05025df8c","relation":"same-reasoning"}]},{"id":"b-0f92d758c53ae5d4a6f20de3","type":"p","scope":"set-matching-real-inputs","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-0f92d758c53ae5d4a6f20de3","content":"当前 LAB20 的第一项预测框在右侧，类别概率更支持第二类；第二项预测在左侧，更支持第一类。真值数组却按左、右保存。如果不考虑内容，直接让槽位 0 学真值 0，就人为制造了位置和类别冲突。","sha256":"8c4a4fcaec0e6c0eb856799d47828d49dcd70cbeddd2d724b59395207d95348e","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":"set-matching-lab20-costs","title":"选读 LAB20：沿真实输入核对交叉匹配","number":8},"sectionTitle":"LAB20 的两个槽位为何应交叉匹配","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-p35","l3-r2-p39"],"knowledgeBlock":"set-matching-real-inputs","formulas":[],"previousBlock":"b-07496f85b77ac1623307e2eb","nextBlock":"b-91a5f7f6a09aaa85af8addb8","prerequisites":[{"id":"b-07496f85b77ac1623307e2eb","url":"#b-07496f85b77ac1623307e2eb","role":"section-title"},{"id":"b-e637ab72577ef03dfac4f03d","url":"#b-e637ab72577ef03dfac4f03d","role":"problem-context"}],"relatedBlocks":[{"id":"b-07496f85b77ac1623307e2eb","url":"#b-07496f85b77ac1623307e2eb","relation":"same-reasoning"},{"id":"b-91a5f7f6a09aaa85af8addb8","url":"#b-91a5f7f6a09aaa85af8addb8","relation":"same-reasoning"},{"id":"b-efb5f1232e4b79525f38a178","url":"#b-efb5f1232e4b79525f38a178","relation":"same-reasoning"},{"id":"b-2d6fed637b1b9ac05025df8c","url":"#b-2d6fed637b1b9ac05025df8c","relation":"same-reasoning"}]},{"id":"b-91a5f7f6a09aaa85af8addb8","type":"p","scope":"set-matching-real-inputs","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-91a5f7f6a09aaa85af8addb8","content":"代码按教学定义 $C_{ij}=-\\log p_i(c_j)+2\\lVert b_i-g_j\\rVert_1+(1-IoU)$ 构造成本，得到约 `[[5.0971,.4509],[.4779,5.6026]]`。对角总代价 10.6997，交叉 .9288，因此最佳分配为槽位 0→真值 1、槽位 1→真值 0。只交换标注或预测数组的排列，最小内容对应仍然不变。","sha256":"666afb23ae32517dfc3373217aa07120f72ac5e338277cf55cef59ec35e29c96","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":"set-matching-lab20-costs","title":"选读 LAB20：沿真实输入核对交叉匹配","number":8},"sectionTitle":"LAB20 的两个槽位为何应交叉匹配","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-p35","l3-r2-p39"],"knowledgeBlock":"set-matching-real-inputs","formulas":[{"latex":"C_{ij}=-\\log p_i(c_j)+2\\lVert b_i-g_j\\rVert_1+(1-IoU)","definitionContext":"当前 LAB20 的第一项预测框在右侧，类别概率更支持第二类；第二项预测在左侧，更支持第一类。真值数组却按左、右保存。如果不考虑内容，直接让槽位 0 学真值 0，就人为制造了位置和类别冲突。\n代码按教学定义 $C_{ij}=-\\log p_i(c_j)+2\\lVert b_i-g_j\\rVert_1+(1-IoU)$ 构造成本，得到约 `[[5.0971,.4509],[.4779,5.6026]]`。对角总代价 10.6997，交叉 .9288，因此最佳分配为槽位 0→真值 1、槽位 1→真值 0。只交换标注或预测数组的排列，最小内容对应仍然不变。\n此处的负对数分类项、IoU 与系数是教学选择；官方 DETR matcher 使用其明确的分类、L1、GIoU 成本，不应声称公式逐项一致。匹配找到最低代价也不意味着检测正确：即使两项预测都很差，算法仍会选一个较便宜分配。检测质量仍要用保留数据与定位/类别标准评价。\n本册另一小算例把 16 像素目标映到不同 stride 的连续跨度，属于尺度读取，与集合匹配不同。导航中把二者标为两项可单独读的短练习，学生无需为找一句机制解释穿过全部实验。"}],"previousBlock":"b-0f92d758c53ae5d4a6f20de3","nextBlock":"b-efb5f1232e4b79525f38a178","prerequisites":[{"id":"b-07496f85b77ac1623307e2eb","url":"#b-07496f85b77ac1623307e2eb","role":"section-title"},{"id":"b-e637ab72577ef03dfac4f03d","url":"#b-e637ab72577ef03dfac4f03d","role":"problem-context"}],"relatedBlocks":[{"id":"b-07496f85b77ac1623307e2eb","url":"#b-07496f85b77ac1623307e2eb","relation":"same-reasoning"},{"id":"b-0f92d758c53ae5d4a6f20de3","url":"#b-0f92d758c53ae5d4a6f20de3","relation":"same-reasoning"},{"id":"b-efb5f1232e4b79525f38a178","url":"#b-efb5f1232e4b79525f38a178","relation":"same-reasoning"},{"id":"b-2d6fed637b1b9ac05025df8c","url":"#b-2d6fed637b1b9ac05025df8c","relation":"same-reasoning"}]},{"id":"b-efb5f1232e4b79525f38a178","type":"p","scope":"set-matching-real-inputs","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-efb5f1232e4b79525f38a178","content":"此处的负对数分类项、IoU 与系数是教学选择；官方 DETR matcher 使用其明确的分类、L1、GIoU 成本，不应声称公式逐项一致。匹配找到最低代价也不意味着检测正确：即使两项预测都很差，算法仍会选一个较便宜分配。检测质量仍要用保留数据与定位/类别标准评价。","sha256":"00cdf3ea105cff573d3b5e30cd194b7c8bb22cc8b37fda235912d4f9ef325925","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":"set-matching-lab20-costs","title":"选读 LAB20：沿真实输入核对交叉匹配","number":8},"sectionTitle":"LAB20 的两个槽位为何应交叉匹配","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-p35","l3-r2-p39"],"knowledgeBlock":"set-matching-real-inputs","formulas":[],"previousBlock":"b-91a5f7f6a09aaa85af8addb8","nextBlock":"b-2d6fed637b1b9ac05025df8c","prerequisites":[{"id":"b-07496f85b77ac1623307e2eb","url":"#b-07496f85b77ac1623307e2eb","role":"section-title"},{"id":"b-e637ab72577ef03dfac4f03d","url":"#b-e637ab72577ef03dfac4f03d","role":"problem-context"}],"relatedBlocks":[{"id":"b-07496f85b77ac1623307e2eb","url":"#b-07496f85b77ac1623307e2eb","relation":"same-reasoning"},{"id":"b-0f92d758c53ae5d4a6f20de3","url":"#b-0f92d758c53ae5d4a6f20de3","relation":"same-reasoning"},{"id":"b-91a5f7f6a09aaa85af8addb8","url":"#b-91a5f7f6a09aaa85af8addb8","relation":"same-reasoning"},{"id":"b-2d6fed637b1b9ac05025df8c","url":"#b-2d6fed637b1b9ac05025df8c","relation":"same-reasoning"}]},{"id":"b-2d6fed637b1b9ac05025df8c","type":"p","scope":"set-matching-real-inputs","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-2d6fed637b1b9ac05025df8c","content":"本册另一小算例把 16 像素目标映到不同 stride 的连续跨度，属于尺度读取，与集合匹配不同。导航中把二者标为两项可单独读的短练习，学生无需为找一句机制解释穿过全部实验。","sha256":"2c901bf98cd806f477e314f0897e9351b3172089a6eab29d4f2569837c1837e6","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":"set-matching-lab20-costs","title":"选读 LAB20：沿真实输入核对交叉匹配","number":8},"sectionTitle":"LAB20 的两个槽位为何应交叉匹配","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-p35","l3-r2-p39"],"knowledgeBlock":"set-matching-real-inputs","formulas":[],"previousBlock":"b-efb5f1232e4b79525f38a178","nextBlock":null,"prerequisites":[{"id":"b-07496f85b77ac1623307e2eb","url":"#b-07496f85b77ac1623307e2eb","role":"section-title"},{"id":"b-e637ab72577ef03dfac4f03d","url":"#b-e637ab72577ef03dfac4f03d","role":"problem-context"}],"relatedBlocks":[{"id":"b-07496f85b77ac1623307e2eb","url":"#b-07496f85b77ac1623307e2eb","relation":"same-reasoning"},{"id":"b-0f92d758c53ae5d4a6f20de3","url":"#b-0f92d758c53ae5d4a6f20de3","relation":"same-reasoning"},{"id":"b-91a5f7f6a09aaa85af8addb8","url":"#b-91a5f7f6a09aaa85af8addb8","relation":"same-reasoning"},{"id":"b-efb5f1232e4b79525f38a178","url":"#b-efb5f1232e4b79525f38a178","relation":"same-reasoning"}]},{"id":"b-e3b9d4c1a5382a7839d845b5","type":"h2","scope":"spatial-practice-selection","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-e3b9d4c1a5382a7839d845b5","content":"## 选择一项完整实验，不按编号机械运行","sha256":"a5362ef585bbbf023132a3de2fef18ea5fe82ca254b8bdf1725d3c2bd4c7e079","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-to-language-and-practice","title":"带着空间接口，怎样进入实验与语言任务？"},"page":{"id":"spatial-practice-selection","title":"选择一项完整实验，不按编号机械运行","number":9},"sectionTitle":"选择一项完整实验，不按编号机械运行","currentQuestion":"先选择需要交付的答案，再选择一个完整学习过程。","sourceSlides":["l3-r2-p25","l3-r2-p35","l3-r2-p38","l3-r2-p48"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-9f21524d34ae60bba14da238","prerequisites":[{"id":"b-9f21524d34ae60bba14da238","url":"#b-9f21524d34ae60bba14da238","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-9f21524d34ae60bba14da238","type":"p","scope":"spatial-practice-selection","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-9f21524d34ae60bba14da238","content":"按科研问题选择定位或像素路径，再用解析练习补齐具体疑问。","sha256":"5f3379499815a12448b7f4656737d34d67dc031279c7ede7996c2b49bd5ed480","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-to-language-and-practice","title":"带着空间接口，怎样进入实验与语言任务？"},"page":{"id":"spatial-practice-selection","title":"选择一项完整实验，不按编号机械运行","number":9},"sectionTitle":"选择一项完整实验，不按编号机械运行","currentQuestion":"先选择需要交付的答案，再选择一个完整学习过程。","sourceSlides":["l3-r2-p25","l3-r2-p35","l3-r2-p38","l3-r2-p48"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e3b9d4c1a5382a7839d845b5","nextBlock":"b-d76aead57f8f31763bbf38dd","prerequisites":[{"id":"b-e3b9d4c1a5382a7839d845b5","url":"#b-e3b9d4c1a5382a7839d845b5","role":"section-title"}],"relatedBlocks":[]},{"id":"b-d76aead57f8f31763bbf38dd","type":"h3","scope":"spatial-study-route","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-d76aead57f8f31763bbf38dd","content":"### 这条主线的课后顺序：先看输出，再做一项完整实验","sha256":"9c1d4e7b42a1c13bd684104ac7d8b248ab71cdd076f2976b642e9cd4d88212ac","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-to-language-and-practice","title":"带着空间接口，怎样进入实验与语言任务？"},"page":{"id":"spatial-practice-selection","title":"选择一项完整实验，不按编号机械运行","number":9},"sectionTitle":"这条主线的课后顺序：先看输出，再做一项完整实验","currentQuestion":"先选择需要交付的答案，再选择一个完整学习过程。","sourceSlides":["l3-r2-p25","l3-r2-p35","l3-r2-p38","l3-r2-p48"],"knowledgeBlock":"spatial-study-route","formulas":[],"previousBlock":"b-9f21524d34ae60bba14da238","nextBlock":"b-3de39f5d21e18cda37f10dc2","prerequisites":[{"id":"b-e3b9d4c1a5382a7839d845b5","url":"#b-e3b9d4c1a5382a7839d845b5","role":"section-title"},{"id":"b-9f21524d34ae60bba14da238","url":"#b-9f21524d34ae60bba14da238","role":"problem-context"}],"relatedBlocks":[{"id":"b-3de39f5d21e18cda37f10dc2","url":"#b-3de39f5d21e18cda37f10dc2","relation":"same-reasoning"},{"id":"b-3de7cc46ed6aec104caec368","url":"#b-3de7cc46ed6aec104caec368","relation":"same-reasoning"},{"id":"b-b088dad2685cf2eec1252f6f","url":"#b-b088dad2685cf2eec1252f6f","relation":"same-reasoning"},{"id":"b-c1e3af93b0c02bc155ada275","url":"#b-c1e3af93b0c02bc155ada275","relation":"same-reasoning"}]},{"id":"b-3de39f5d21e18cda37f10dc2","type":"p","scope":"spatial-study-route","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-3de39f5d21e18cda37f10dc2","content":"不必按 Notebook 编号把所有内容从头运行。若刚听完课堂，先用 P25 的同一底图区分类别、框、语义与实例，再用 P41–45 的两个算例区分读取错位与表示碰撞。它们构成理解空间任务所需的最短连续线。","sha256":"0a2ecad891c31a62ee9b02e86cdba8178bd3d109c92dba9a0311468b72ee9cd0","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-to-language-and-practice","title":"带着空间接口，怎样进入实验与语言任务？"},"page":{"id":"spatial-practice-selection","title":"选择一项完整实验，不按编号机械运行","number":9},"sectionTitle":"这条主线的课后顺序：先看输出，再做一项完整实验","currentQuestion":"先选择需要交付的答案，再选择一个完整学习过程。","sourceSlides":["l3-r2-p25","l3-r2-p35","l3-r2-p38","l3-r2-p48"],"knowledgeBlock":"spatial-study-route","formulas":[],"previousBlock":"b-d76aead57f8f31763bbf38dd","nextBlock":"b-3de7cc46ed6aec104caec368","prerequisites":[{"id":"b-d76aead57f8f31763bbf38dd","url":"#b-d76aead57f8f31763bbf38dd","role":"section-title"},{"id":"b-9f21524d34ae60bba14da238","url":"#b-9f21524d34ae60bba14da238","role":"problem-context"}],"relatedBlocks":[{"id":"b-d76aead57f8f31763bbf38dd","url":"#b-d76aead57f8f31763bbf38dd","relation":"same-reasoning"},{"id":"b-3de7cc46ed6aec104caec368","url":"#b-3de7cc46ed6aec104caec368","relation":"same-reasoning"},{"id":"b-b088dad2685cf2eec1252f6f","url":"#b-b088dad2685cf2eec1252f6f","relation":"same-reasoning"},{"id":"b-c1e3af93b0c02bc155ada275","url":"#b-c1e3af93b0c02bc155ada275","relation":"same-reasoning"}]},{"id":"b-3de7cc46ed6aec104caec368","type":"p","scope":"spatial-study-route","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-3de7cc46ed6aec104caec368","content":"随后只选一项完整学习过程：关心对象位置就读 LAB10，沿图像、目标、匹配、分类/位置损失、预测与错误图走完；关心面积或边界就读 LAB11，比较 RGB 基线与 FCN、查看逐类 IoU 与窄缝错误。无需为了“完整”同时训练两本，也不要把只跑通环境当成理解模型。","sha256":"6c778237b86f2e11d4c30d82b852d6200f4b10ccec5ed4c1beb3aad463157942","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-to-language-and-practice","title":"带着空间接口，怎样进入实验与语言任务？"},"page":{"id":"spatial-practice-selection","title":"选择一项完整实验，不按编号机械运行","number":9},"sectionTitle":"这条主线的课后顺序：先看输出，再做一项完整实验","currentQuestion":"先选择需要交付的答案，再选择一个完整学习过程。","sourceSlides":["l3-r2-p25","l3-r2-p35","l3-r2-p38","l3-r2-p48"],"knowledgeBlock":"spatial-study-route","formulas":[],"previousBlock":"b-3de39f5d21e18cda37f10dc2","nextBlock":"b-b088dad2685cf2eec1252f6f","prerequisites":[{"id":"b-d76aead57f8f31763bbf38dd","url":"#b-d76aead57f8f31763bbf38dd","role":"section-title"},{"id":"b-9f21524d34ae60bba14da238","url":"#b-9f21524d34ae60bba14da238","role":"problem-context"}],"relatedBlocks":[{"id":"b-d76aead57f8f31763bbf38dd","url":"#b-d76aead57f8f31763bbf38dd","relation":"same-reasoning"},{"id":"b-3de39f5d21e18cda37f10dc2","url":"#b-3de39f5d21e18cda37f10dc2","relation":"same-reasoning"},{"id":"b-b088dad2685cf2eec1252f6f","url":"#b-b088dad2685cf2eec1252f6f","relation":"same-reasoning"},{"id":"b-c1e3af93b0c02bc155ada275","url":"#b-c1e3af93b0c02bc155ada275","relation":"same-reasoning"}]},{"id":"b-b088dad2685cf2eec1252f6f","type":"p","scope":"spatial-study-route","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-b088dad2685cf2eec1252f6f","content":"LAB09 是检测公式的工作台，遇到匹配、NMS 或 AP 不清时按问题查；LAB03 是同源场景与空间读数的轻量练习。LAB16 的颜色连通域和 LAB20 的集合匹配放在选读，分别解释提示约束和对象排列，不与两项主实验争夺主线。","sha256":"5bd92268b823ca1189663b03a9aab5ad528b77f1824da7041e68e16660f8a280","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-to-language-and-practice","title":"带着空间接口，怎样进入实验与语言任务？"},"page":{"id":"spatial-practice-selection","title":"选择一项完整实验，不按编号机械运行","number":9},"sectionTitle":"这条主线的课后顺序：先看输出，再做一项完整实验","currentQuestion":"先选择需要交付的答案，再选择一个完整学习过程。","sourceSlides":["l3-r2-p25","l3-r2-p35","l3-r2-p38","l3-r2-p48"],"knowledgeBlock":"spatial-study-route","formulas":[],"previousBlock":"b-3de7cc46ed6aec104caec368","nextBlock":"b-c1e3af93b0c02bc155ada275","prerequisites":[{"id":"b-d76aead57f8f31763bbf38dd","url":"#b-d76aead57f8f31763bbf38dd","role":"section-title"},{"id":"b-9f21524d34ae60bba14da238","url":"#b-9f21524d34ae60bba14da238","role":"problem-context"}],"relatedBlocks":[{"id":"b-d76aead57f8f31763bbf38dd","url":"#b-d76aead57f8f31763bbf38dd","relation":"same-reasoning"},{"id":"b-3de39f5d21e18cda37f10dc2","url":"#b-3de39f5d21e18cda37f10dc2","relation":"same-reasoning"},{"id":"b-3de7cc46ed6aec104caec368","url":"#b-3de7cc46ed6aec104caec368","relation":"same-reasoning"},{"id":"b-c1e3af93b0c02bc155ada275","url":"#b-c1e3af93b0c02bc155ada275","relation":"same-reasoning"}]},{"id":"b-c1e3af93b0c02bc155ada275","type":"p","scope":"spatial-study-route","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-c1e3af93b0c02bc155ada275","content":"CLS02 提供区域到掩码的连续导读；CLS07、CLS08 分别展开检测公式与实例输出。各处按教师定稿名称和稳定 ID 回链。每个解释块有独立 URL，可直接带着具体问题交给 AI；链接中的正文先提供推理，再连接代码与原论文。","sha256":"7dc68c2032c902245b20b5e27e414e7e048471fa6a74d58eab62ea8327fd9a26","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-to-language-and-practice","title":"带着空间接口，怎样进入实验与语言任务？"},"page":{"id":"spatial-practice-selection","title":"选择一项完整实验，不按编号机械运行","number":9},"sectionTitle":"这条主线的课后顺序：先看输出，再做一项完整实验","currentQuestion":"先选择需要交付的答案，再选择一个完整学习过程。","sourceSlides":["l3-r2-p25","l3-r2-p35","l3-r2-p38","l3-r2-p48"],"knowledgeBlock":"spatial-study-route","formulas":[],"previousBlock":"b-b088dad2685cf2eec1252f6f","nextBlock":"b-30c7edf0ddb51751ac361073","prerequisites":[{"id":"b-d76aead57f8f31763bbf38dd","url":"#b-d76aead57f8f31763bbf38dd","role":"section-title"},{"id":"b-9f21524d34ae60bba14da238","url":"#b-9f21524d34ae60bba14da238","role":"problem-context"}],"relatedBlocks":[{"id":"b-d76aead57f8f31763bbf38dd","url":"#b-d76aead57f8f31763bbf38dd","relation":"same-reasoning"},{"id":"b-3de39f5d21e18cda37f10dc2","url":"#b-3de39f5d21e18cda37f10dc2","relation":"same-reasoning"},{"id":"b-3de7cc46ed6aec104caec368","url":"#b-3de7cc46ed6aec104caec368","relation":"same-reasoning"},{"id":"b-b088dad2685cf2eec1252f6f","url":"#b-b088dad2685cf2eec1252f6f","relation":"same-reasoning"}]},{"id":"b-30c7edf0ddb51751ac361073","type":"h2","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-30c7edf0ddb51751ac361073","content":"## 沿这个问题继续","sha256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-to-language-and-practice","title":"带着空间接口，怎样进入实验与语言任务？"},"page":{"id":"spatial-practice-selection","title":"选择一项完整实验，不按编号机械运行","number":9},"sectionTitle":"选择一项完整实验，不按编号机械运行","currentQuestion":"先选择需要交付的答案，再选择一个完整学习过程。","sourceSlides":["l3-r2-p25","l3-r2-p35","l3-r2-p38","l3-r2-p48"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c1e3af93b0c02bc155ada275","nextBlock":"b-a0cf6fcaa14744440d0133e4","prerequisites":[{"id":"b-d76aead57f8f31763bbf38dd","url":"#b-d76aead57f8f31763bbf38dd","role":"section-title"},{"id":"b-9f21524d34ae60bba14da238","url":"#b-9f21524d34ae60bba14da238","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-a0cf6fcaa14744440d0133e4","type":"p","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-a0cf6fcaa14744440d0133e4","content":"先沿 [CLS07](/course/notebooks/586e66e049764158a50747e69c219312.html) 定位检测几何与评价疑问，再选 [LAB09](/course/notebooks/90d213b4e1e04c7e889ee6fc042b2358.html) 做解析练习，或沿 [LAB10](/course/notebooks/a14237bfa9dd413d959ab4e458e11f42.html) 走完一次完整检测学习。其余入口见[本部分更多学习路径](/course/blocks/spatial-study-route.html)。","sha256":"8f2c63167e56ee768534601ac56a55e07567c8792a5deb33e7da197b37f632c9","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-to-language-and-practice","title":"带着空间接口，怎样进入实验与语言任务？"},"page":{"id":"spatial-practice-selection","title":"选择一项完整实验，不按编号机械运行","number":9},"sectionTitle":"选择一项完整实验，不按编号机械运行","currentQuestion":"先选择需要交付的答案，再选择一个完整学习过程。","sourceSlides":["l3-r2-p25","l3-r2-p35","l3-r2-p38","l3-r2-p48"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-30c7edf0ddb51751ac361073","nextBlock":"b-271a42cf91f68e59411c2a70","prerequisites":[{"id":"b-30c7edf0ddb51751ac361073","url":"#b-30c7edf0ddb51751ac361073","role":"section-title"},{"id":"b-9f21524d34ae60bba14da238","url":"#b-9f21524d34ae60bba14da238","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-271a42cf91f68e59411c2a70","type":"h4","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p35.html#b-271a42cf91f68e59411c2a70","content":"#### 按问题回查的讲解与短练习（6）","sha256":"5d4f2896329b48980886361effc445db45704337dde3d7e7242461994c153117","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"spatial-to-language-and-practice","title":"带着空间接口，怎样进入实验与语言任务？"},"page":{"id":"spatial-practice-selection","title":"选择一项完整实验，不按编号机械运行","number":9},"sectionTitle":"选择一项完整实验，不按编号机械运行","currentQuestion":"先选择需要交付的答案，再选择一个完整学习过程。","sourceSlides":["l3-r2-p25","l3-r2-p35","l3-r2-p38","l3-r2-p48"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-a0cf6fcaa14744440d0133e4","nextBlock":null,"prerequisites":[{"id":"b-30c7edf0ddb51751ac361073","url":"#b-30c7edf0ddb51751ac361073","role":"section-title"},{"id":"b-9f21524d34ae60bba14da238","url":"#b-9f21524d34ae60bba14da238","role":"problem-context"}],"relatedBlocks":[]}]}