{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"CLS 08 · 密集预测、实例与集合输出","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html","markdown":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.md","pages":[{"id":"reading-opening","title":"CLS 08 · 密集预测、实例与集合输出","unit":"CLS 08 · 密集预测、实例与集合输出","unitId":"cls09-密集预测实例与集合输出","chapter":"course","number":1},{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","unit":"CLS 08 · 密集预测、实例与集合输出","unitId":"cls09-密集预测实例与集合输出","chapter":"course","number":2},{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","unit":"CLS 08 · 密集预测、实例与集合输出","unitId":"cls09-密集预测实例与集合输出","chapter":"course","number":3},{"id":"reading-block-59e4b23da62d4b709cc87f278811ec2f","title":"对齐的正面实证","unit":"CLS 08 · 密集预测、实例与集合输出","unitId":"cls09-密集预测实例与集合输出","chapter":"course","number":4}],"aliases":[{"oldId":"b-6629d08e246a615ce015dcf9","targetId":"b-6629d08e246a615ce015dcf9","operation":"historical-explanation","sourceScope":"block-8413a612fc314b05806c00e9cf00fb7a"},{"oldId":"b-4bfc2e517d2ac8858e219062","targetId":"b-4bfc2e517d2ac8858e219062","operation":"historical-explanation","sourceScope":"block-8413a612fc314b05806c00e9cf00fb7a"},{"oldId":"b-6cdbf66916d7ed71b3374fcd","targetId":"b-6cdbf66916d7ed71b3374fcd","operation":"historical-explanation","sourceScope":"block-8413a612fc314b05806c00e9cf00fb7a"},{"oldId":"b-e830b8bc77d57b92862563a1","targetId":"b-e830b8bc77d57b92862563a1","operation":"historical-explanation","sourceScope":"block-8413a612fc314b05806c00e9cf00fb7a"},{"oldId":"b-256c7bedef66c034c7868751","targetId":"b-256c7bedef66c034c7868751","page":"reading-opening","operation":"move","beforeSHA256":"e4f817838c06f025f95e906deb158142e9364a9ad54a95e0fa2fdf2414eeb1d4","afterSHA256":"e4f817838c06f025f95e906deb158142e9364a9ad54a95e0fa2fdf2414eeb1d4","reason":"Same source body in authored semantic page"},{"oldId":"b-ad4b26c04754f893d870d4c3","targetId":"b-ad4b26c04754f893d870d4c3","page":"reading-opening","operation":"move","beforeSHA256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","afterSHA256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","reason":"Same source body in authored semantic page"},{"oldId":"b-4cb0c35f5239fbcfba4d674a","targetId":"b-4cb0c35f5239fbcfba4d674a","page":"reading-opening","operation":"move","beforeSHA256":"70eec56f79b131d4695a1c420bc695a02ef3863727481994de8f7ed4bd673dcc","afterSHA256":"70eec56f79b131d4695a1c420bc695a02ef3863727481994de8f7ed4bd673dcc","reason":"Same source body in authored semantic page"},{"oldId":"b-3feef70df269f2b7ae546bef","targetId":"b-3feef70df269f2b7ae546bef","page":"reading-opening","operation":"move","beforeSHA256":"f22152fd55dcdb0090b5653d37145694fd19388d31e10226549464fe8b483578","afterSHA256":"f22152fd55dcdb0090b5653d37145694fd19388d31e10226549464fe8b483578","reason":"Same source body in authored semantic page"},{"oldId":"b-7bf8697770ccfefdf868ec31","targetId":"b-7bf8697770ccfefdf868ec31","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"30ab9f304366670818536cf73c98c66aa814e8e31ecdbffcf19a4d6e7e7f22de","afterSHA256":"30ab9f304366670818536cf73c98c66aa814e8e31ecdbffcf19a4d6e7e7f22de","reason":"Same source body in authored semantic page"},{"oldId":"b-7824015c8fa9007ea0c7f8e4","targetId":"b-7824015c8fa9007ea0c7f8e4","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"72798cbfc37e71e8b9dbbcfe3cfb8a7b0fc2d74531b251cdcb3e0c9989b9c37d","afterSHA256":"72798cbfc37e71e8b9dbbcfe3cfb8a7b0fc2d74531b251cdcb3e0c9989b9c37d","reason":"Same source body in authored semantic page"},{"oldId":"b-e9bb6e3b3e8ff1a4f99a460d","targetId":"b-e9bb6e3b3e8ff1a4f99a460d","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"93743e4257248bf9219358b363aae08572036e7ccafb37b6917bd45eeacfbe96","afterSHA256":"93743e4257248bf9219358b363aae08572036e7ccafb37b6917bd45eeacfbe96","reason":"Same source body in authored semantic page"},{"oldId":"b-c06dd4bf7393cdb19ca92b7b","targetId":"b-c06dd4bf7393cdb19ca92b7b","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"ad5c18536ebaaed1683a9cbbc42e20e2f10a94f3ad582f81270f89de1719a98f","afterSHA256":"ad5c18536ebaaed1683a9cbbc42e20e2f10a94f3ad582f81270f89de1719a98f","reason":"Same source body in authored semantic page"},{"oldId":"b-9f9fa2f87c514087cde4da29","targetId":"b-9f9fa2f87c514087cde4da29","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"5409b0105bb5ec9dc3241c580e59c54772c6e8b9d8dbcfe5caf959343ea238a6","afterSHA256":"5409b0105bb5ec9dc3241c580e59c54772c6e8b9d8dbcfe5caf959343ea238a6","reason":"Same source body in authored semantic page"},{"oldId":"b-b07b8d541342c12da8f0a3f7","targetId":"b-b07b8d541342c12da8f0a3f7","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"77f6d53302f1be4d6d3ee57e02b21459f56e0aa208d6bbbd81accd9d8b377d53","afterSHA256":"77f6d53302f1be4d6d3ee57e02b21459f56e0aa208d6bbbd81accd9d8b377d53","reason":"Same source body in authored semantic page"},{"oldId":"b-a6d93d405af42d1c53096d4c","targetId":"b-a6d93d405af42d1c53096d4c","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"3e56a076a7b9a48f3f8c0e4a1ec3518b7589369f6099f713d3ab613354f864c0","afterSHA256":"3e56a076a7b9a48f3f8c0e4a1ec3518b7589369f6099f713d3ab613354f864c0","reason":"Same source body in authored semantic page"},{"oldId":"b-143c0b332ba51e83ccc3baef","targetId":"b-143c0b332ba51e83ccc3baef","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"ad790b051493beed5374493af79f9514dc8aa5e0b22af3b3dc4d840d31e23fe3","afterSHA256":"ad790b051493beed5374493af79f9514dc8aa5e0b22af3b3dc4d840d31e23fe3","reason":"Same source body in authored semantic page"},{"oldId":"b-65acb3f8c752f24c2f082610","targetId":"b-65acb3f8c752f24c2f082610","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"be0aaf3d04041ca7d223159fab6db6c9adaac7193b5a1788872d4d17ef5a188b","afterSHA256":"be0aaf3d04041ca7d223159fab6db6c9adaac7193b5a1788872d4d17ef5a188b","reason":"Same source body in authored semantic page"},{"oldId":"b-bc6e32b035fa71a395b40263","targetId":"b-bc6e32b035fa71a395b40263","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"a9149806afd622bccc16835a043e5f4608c2cf76dcabe0e7596a0cc5c2bf8c51","afterSHA256":"a9149806afd622bccc16835a043e5f4608c2cf76dcabe0e7596a0cc5c2bf8c51","reason":"Same source body in authored semantic page"},{"oldId":"b-8fdcd1281b281fe4293dc8d1","targetId":"b-8fdcd1281b281fe4293dc8d1","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"241bf275da03aa8d9391b79ee6cd44b57a829f6baa38512357bc65f453f808df","afterSHA256":"241bf275da03aa8d9391b79ee6cd44b57a829f6baa38512357bc65f453f808df","reason":"Same source body in authored semantic page"},{"oldId":"b-4fa0186098005215f085fe0c","targetId":"b-4fa0186098005215f085fe0c","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"cf7413ba356779e875fe91f4f34c7c315f7f7b7ce42e19b44ffd1dcb30d0d58b","afterSHA256":"cf7413ba356779e875fe91f4f34c7c315f7f7b7ce42e19b44ffd1dcb30d0d58b","reason":"Same source body in authored semantic page"},{"oldId":"b-6d725477d10319e2c54bdad8","targetId":"b-6d725477d10319e2c54bdad8","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"041babd888393a13f260ef325bcc31c4c9cd1e4f7df2491e324ddafb2b56cd23","afterSHA256":"041babd888393a13f260ef325bcc31c4c9cd1e4f7df2491e324ddafb2b56cd23","reason":"Same source body in authored semantic page"},{"oldId":"b-a43c41cc687b47b717ec77c9","targetId":"b-a43c41cc687b47b717ec77c9","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"c315ee35160727db2c7bb7bef724e45266800f78ea5d2880c7d3101479ad3b43","afterSHA256":"c315ee35160727db2c7bb7bef724e45266800f78ea5d2880c7d3101479ad3b43","reason":"Same source body in authored semantic page"},{"oldId":"b-ea8636c8daedf08fdeff6409","targetId":"b-ea8636c8daedf08fdeff6409","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"9d6d23d8e3c027ed4f0115e54b8a2a363b6e2f8a619951c7433c03f79bcd0992","afterSHA256":"9d6d23d8e3c027ed4f0115e54b8a2a363b6e2f8a619951c7433c03f79bcd0992","reason":"Same source body in authored semantic page"},{"oldId":"b-d7e7e9870d55b802a8b461d1","targetId":"b-d7e7e9870d55b802a8b461d1","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"ae393bf00717f2b8ab32d920a04af01319f2e21853fbca2d696f4bb48f770610","afterSHA256":"ae393bf00717f2b8ab32d920a04af01319f2e21853fbca2d696f4bb48f770610","reason":"Same source body in authored semantic page"},{"oldId":"b-47a7abca10abbc6fc0131e80","targetId":"b-47a7abca10abbc6fc0131e80","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"35d7aa5a6598e52321e9309a0916b06dc03cd25c5ad4df6383272b464361299a","afterSHA256":"35d7aa5a6598e52321e9309a0916b06dc03cd25c5ad4df6383272b464361299a","reason":"Same source body in authored semantic page"},{"oldId":"b-58c643129233b5bf3fcdf363","targetId":"b-58c643129233b5bf3fcdf363","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"cdf5e1ee80a89ea700fef3561ecdfa2d3713ae806370d2924e5897e2a5078529","afterSHA256":"cdf5e1ee80a89ea700fef3561ecdfa2d3713ae806370d2924e5897e2a5078529","reason":"Same source body in authored semantic page"},{"oldId":"b-979030e464e9fa7cadcd3fa4","targetId":"b-979030e464e9fa7cadcd3fa4","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"d8c17ff28b5b0d5b1671238ddcef7a1c477eeded2b48b525dadaddb86200d640","afterSHA256":"d8c17ff28b5b0d5b1671238ddcef7a1c477eeded2b48b525dadaddb86200d640","reason":"Same source body in authored semantic page"},{"oldId":"b-d367e0432b7c153aa011785b","targetId":"b-d367e0432b7c153aa011785b","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"13df4d569b221f8bc60c5de960a783d1972c800851f3e5ebd251ca5ca31bb27c","afterSHA256":"13df4d569b221f8bc60c5de960a783d1972c800851f3e5ebd251ca5ca31bb27c","reason":"Same source body in authored semantic page"},{"oldId":"b-27cdb17f336635825aa19e6f","targetId":"b-27cdb17f336635825aa19e6f","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"rewrite","beforeSHA256":"5530957dfc389313134c1ae2e8fbba7cb2ec21f5033903c548698fd1182a0ee5","afterSHA256":"d12b72281bce9d27865a10e904405cca1c2c915b9910c63934ad6f7d4439e49a","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-6a5d378fb1f084605094a70a","targetId":"b-6a5d378fb1f084605094a70a","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"4c2978607bb7b437fdad1528eb065e99e5968fdc50592f9201ed43dcf3ded1a3","afterSHA256":"4c2978607bb7b437fdad1528eb065e99e5968fdc50592f9201ed43dcf3ded1a3","reason":"Same source body in authored semantic page"},{"oldId":"b-b0d1ded46e6bebb316c4efe2","targetId":"b-b0d1ded46e6bebb316c4efe2","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"939045e192e80393bc29ee5638c03a5341285cb4e18440f27403c32464cc6f54","afterSHA256":"939045e192e80393bc29ee5638c03a5341285cb4e18440f27403c32464cc6f54","reason":"Same source body in authored semantic page"},{"oldId":"b-ec127c6207c04d9876a76d11","targetId":"b-ec127c6207c04d9876a76d11","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"715899b38a78b3bf4f5526299aecfd4f7c5178ff28d2193adb58e1166be1ad9d","afterSHA256":"715899b38a78b3bf4f5526299aecfd4f7c5178ff28d2193adb58e1166be1ad9d","reason":"Same source body in authored semantic page"},{"oldId":"b-7987bf989df1a4f31935446a","targetId":"b-7987bf989df1a4f31935446a","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"rewrite","beforeSHA256":"1bdd695c84d8c10d3518a5f111bf260a1bf6eb69d2b1d291cf172813e2d66865","afterSHA256":"506f6e140199a9480864e32072bf7dc411386e1f8072f90aaf04d4a7616bf87e","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-a44a66cfeed574e75874bced","targetId":"b-a44a66cfeed574e75874bced","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"596f845a43d7a997861347a6bbdfe900ef177b21daf6e5f5bc7a191ee6b9e7f8","afterSHA256":"596f845a43d7a997861347a6bbdfe900ef177b21daf6e5f5bc7a191ee6b9e7f8","reason":"Same source body in authored semantic page"},{"oldId":"b-dfb5a10e7e678b16d9e2eb51","targetId":"b-dfb5a10e7e678b16d9e2eb51","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"3692ed029bf31213ee6ddce712d05db3ed57c54d18122d873965128ee09e9c12","afterSHA256":"f56f575033bcab034622fd5bcf423ca70f6f99a0671bfa0c2557bc4ade38d310","reason":"Same source body in authored semantic page"},{"oldId":"b-1fa6ef11953df3de99116beb","targetId":"b-1fa6ef11953df3de99116beb","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"0bc30a0c85f509f96425ef6d503b1f4dda54adc0ee2017a6d11f91fd6398d7a1","afterSHA256":"6e59fe129adbd7f631cd6517b8715f8e88361fc81316790d51a971ff6a7ac9d1","reason":"Same source body in authored semantic page"},{"oldId":"b-5c19a3719eb906055e61116c","targetId":"b-5c19a3719eb906055e61116c","page":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","operation":"move","beforeSHA256":"239d41a1854eaf46c28660b8f1fafbd9e2dd5243ba3e50949f8b9548fb79b67f","afterSHA256":"239d41a1854eaf46c28660b8f1fafbd9e2dd5243ba3e50949f8b9548fb79b67f","reason":"Same source body in authored semantic page"},{"oldId":"b-88083cc9f29886ff67d15c35","targetId":"b-88083cc9f29886ff67d15c35","page":"reading-block-8413a612fc314b05806c00e9cf00fb7a","operation":"move","beforeSHA256":"e4fe0386a7ec4009e0b726ed0380c146b413771e54cc9b67fa42f3c17ddc9e69","afterSHA256":"e4fe0386a7ec4009e0b726ed0380c146b413771e54cc9b67fa42f3c17ddc9e69","reason":"Same source body in authored semantic page"},{"oldId":"b-e4182d7601f80e13d00c0c32","targetId":"b-e4182d7601f80e13d00c0c32","page":"reading-block-8413a612fc314b05806c00e9cf00fb7a","operation":"move","beforeSHA256":"62cd3c3e7472a4976a8c101608c68d04df6eb55e6ed8b1cdfe8b808e43846633","afterSHA256":"62cd3c3e7472a4976a8c101608c68d04df6eb55e6ed8b1cdfe8b808e43846633","reason":"Same source body in authored semantic page"},{"oldId":"b-59d22d3f12d4a46fa39d9236","targetId":"b-59d22d3f12d4a46fa39d9236","page":"reading-block-8413a612fc314b05806c00e9cf00fb7a","operation":"move","beforeSHA256":"b77c69f411f0cec66906efa93953096b72fee0517dd04ceb2eb4f2c14fca2c6c","afterSHA256":"b77c69f411f0cec66906efa93953096b72fee0517dd04ceb2eb4f2c14fca2c6c","reason":"Same source body in authored semantic page"},{"oldId":"b-e791c92f48d99504eec4e562","targetId":"b-e791c92f48d99504eec4e562","page":"reading-block-8413a612fc314b05806c00e9cf00fb7a","operation":"move","beforeSHA256":"df469b42a008cfb7ecd8a88fd7f7836c4bc0779d31028b4938705dc288908ff6","afterSHA256":"df469b42a008cfb7ecd8a88fd7f7836c4bc0779d31028b4938705dc288908ff6","reason":"Same source body in authored semantic page"},{"oldId":"b-763d584a8643606e94939c2b","targetId":"b-763d584a8643606e94939c2b","page":"reading-block-8413a612fc314b05806c00e9cf00fb7a","operation":"move","beforeSHA256":"ea717067584e344418a930c49862e5bea35e54c3e75453a5dc35674aa7112bb5","afterSHA256":"ea717067584e344418a930c49862e5bea35e54c3e75453a5dc35674aa7112bb5","reason":"Same source body in authored semantic page"},{"oldId":"b-9b2fdf37cd22b7d4e05509d9","targetId":"b-9b2fdf37cd22b7d4e05509d9","page":"reading-block-8413a612fc314b05806c00e9cf00fb7a","operation":"move","beforeSHA256":"fdde3520bab89e32b78dca4f14b82d41aa943c01e54d2b8a789f56ae0adbb053","afterSHA256":"fdde3520bab89e32b78dca4f14b82d41aa943c01e54d2b8a789f56ae0adbb053","reason":"Same source body in authored semantic page"},{"oldId":"b-5bdd749d5bc44be721f1010a","targetId":"b-5bdd749d5bc44be721f1010a","page":"reading-block-8413a612fc314b05806c00e9cf00fb7a","operation":"move","beforeSHA256":"e9fee9ad6cac4094a8f46cb3106e8782d3fee4eecd1c1f2207ca52af17c2dec0","afterSHA256":"e9fee9ad6cac4094a8f46cb3106e8782d3fee4eecd1c1f2207ca52af17c2dec0","reason":"Same source body in authored semantic page"},{"oldId":"b-84e1878787815e7595dda972","targetId":"b-84e1878787815e7595dda972","page":"reading-block-8413a612fc314b05806c00e9cf00fb7a","operation":"rewrite","beforeSHA256":"eb9ce84712b6c0b764c42fda389bb02c8469bedac54010965c3b3f790e46d4e0","afterSHA256":"e2deae2a61518dc33f64cf5fcf53f9feb4901c2f5cf511ac72edb945dd9ab022","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-6eabbd70e912c0ae34c10cb3","targetId":"b-6eabbd70e912c0ae34c10cb3","page":"reading-block-8413a612fc314b05806c00e9cf00fb7a","operation":"move","beforeSHA256":"b2da5a7d8caaa632c7ae9f9f6ce0a06fe464eecb40e6472dfe0da761208eb634","afterSHA256":"b2da5a7d8caaa632c7ae9f9f6ce0a06fe464eecb40e6472dfe0da761208eb634","reason":"Same source body in authored semantic page"},{"oldId":"b-112423ff84ed57e9e701e5b3","targetId":"b-112423ff84ed57e9e701e5b3","page":"reading-block-8413a612fc314b05806c00e9cf00fb7a","operation":"move","beforeSHA256":"4f70352dd5540e712df3eec48a1d8a256a043e13993dd8de8b43b7b9f446b8b8","afterSHA256":"4f70352dd5540e712df3eec48a1d8a256a043e13993dd8de8b43b7b9f446b8b8","reason":"Same source body in authored semantic page"},{"oldId":"b-fcfea371257d89215d96ac41","targetId":"b-fcfea371257d89215d96ac41","page":"reading-block-8413a612fc314b05806c00e9cf00fb7a","operation":"move","beforeSHA256":"acad29e92db22a1ad00a309903470a4328f8c18cfdf4fac20c3446621f0f6335","afterSHA256":"acad29e92db22a1ad00a309903470a4328f8c18cfdf4fac20c3446621f0f6335","reason":"Same source body in authored semantic page"},{"oldId":"b-7166a465c2c5e0f2b4734595","targetId":"b-7166a465c2c5e0f2b4734595","page":"reading-block-8413a612fc314b05806c00e9cf00fb7a","operation":"move","beforeSHA256":"6d41a42fe14ef8b81dbbf4c540876335e4e04e4ecf75f45c1c416ef808b7f693","afterSHA256":"6d41a42fe14ef8b81dbbf4c540876335e4e04e4ecf75f45c1c416ef808b7f693","reason":"Same source body in authored semantic page"},{"oldId":"b-acd83acacdc33fad32e43c4a","targetId":"b-acd83acacdc33fad32e43c4a","page":"reading-block-8413a612fc314b05806c00e9cf00fb7a","operation":"move","beforeSHA256":"6ccfb455e3b10fd39f061e5a79711a72b3121ff01414683996e829721f063b92","afterSHA256":"6ccfb455e3b10fd39f061e5a79711a72b3121ff01414683996e829721f063b92","reason":"Same source body in authored semantic page"},{"oldId":"b-183259a5cc28aa462c2bff45","targetId":"b-183259a5cc28aa462c2bff45","page":"reading-block-8413a612fc314b05806c00e9cf00fb7a","operation":"move","beforeSHA256":"a27c6258b1428ede911a5cd9e234dac0995bdd9caa159cef345082b9ebf07a51","afterSHA256":"a27c6258b1428ede911a5cd9e234dac0995bdd9caa159cef345082b9ebf07a51","reason":"Same source body in authored semantic page"},{"oldId":"b-39c6b324fcb45dc258f2d5ec","targetId":"b-39c6b324fcb45dc258f2d5ec","page":"reading-block-59e4b23da62d4b709cc87f278811ec2f","operation":"move","beforeSHA256":"44c753d27f6a83b57ddfb3587ff00d4a897cf8a6fa52e22d06bc9714f5b85014","afterSHA256":"44c753d27f6a83b57ddfb3587ff00d4a897cf8a6fa52e22d06bc9714f5b85014","reason":"Same source body in authored semantic page"},{"oldId":"b-9f136b9fb2d77584285a5379","targetId":"b-9f136b9fb2d77584285a5379","page":"reading-block-59e4b23da62d4b709cc87f278811ec2f","operation":"move","beforeSHA256":"3bb8d279d8595e2246936d624e3b4edea6ceca91fece02cd289fada770a219f4","afterSHA256":"3bb8d279d8595e2246936d624e3b4edea6ceca91fece02cd289fada770a219f4","reason":"Same source body in authored semantic page"},{"oldId":"b-1aee8deba52655a72cd8d405","targetId":"b-1aee8deba52655a72cd8d405","page":"reading-block-59e4b23da62d4b709cc87f278811ec2f","operation":"rewrite","beforeSHA256":"0bfedb1ed97588d37565883224ab357ea9af5995f6404d1f3d1042d261fbdfd8","afterSHA256":"f59df6ce113db0faf2717e5b8440d85ef1f069b2e8b327b430459828cfd41119","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-3546af0154b89215a19aa729","targetId":"b-3546af0154b89215a19aa729","page":"reading-block-59e4b23da62d4b709cc87f278811ec2f","operation":"move","beforeSHA256":"4cdba8cde1b8262a042dd6aa8cda6abe7fd5a4e8cb6f224f5bb95e0f3611cd8d","afterSHA256":"4cdba8cde1b8262a042dd6aa8cda6abe7fd5a4e8cb6f224f5bb95e0f3611cd8d","reason":"Same source body in authored semantic page"},{"oldId":"b-926032e76429a0df5a954b76","targetId":"b-926032e76429a0df5a954b76","page":"reading-block-59e4b23da62d4b709cc87f278811ec2f","operation":"move","beforeSHA256":"a75c59c916a80252869b500277c7120d64fcb5056b16b7893076a6f49e7c3632","afterSHA256":"a75c59c916a80252869b500277c7120d64fcb5056b16b7893076a6f49e7c3632","reason":"Same source body in authored semantic page"},{"oldId":"b-a1af8ade7f5836656416e1e2","targetId":"b-a1af8ade7f5836656416e1e2","page":"reading-block-59e4b23da62d4b709cc87f278811ec2f","operation":"move","beforeSHA256":"1362761b98311ffab60c114fa4dc5d8cdf666b37458ced4e4b9d4483229b8e7e","afterSHA256":"1362761b98311ffab60c114fa4dc5d8cdf666b37458ced4e4b9d4483229b8e7e","reason":"Same source body in authored semantic page"},{"oldId":"block-6ca34c15abaf4aeb9900af72759ebc28--一张类别图怎样成为可训练的输出","targetId":"b-b07b8d541342c12da8f0a3f7","operation":"retain-source-identity"},{"oldId":"block-6ca34c15abaf4aeb9900af72759ebc28--上采样能增加格子不能单凭格子恢复被删掉的区别","targetId":"b-8fdcd1281b281fe4293dc8d1","operation":"retain-source-identity"},{"oldId":"block-6ca34c15abaf4aeb9900af72759ebc28--一个-88-反例怎样澄清四种指标","targetId":"b-47a7abca10abbc6fc0131e80","operation":"retain-source-identity"},{"oldId":"block-8413a612fc314b05806c00e9cf00fb7a--从类别像素走向逐个对象","targetId":"b-88083cc9f29886ff67d15c35","operation":"retain-source-identity"},{"oldId":"block-59e4b23da62d4b709cc87f278811ec2f--对齐的正面实证","targetId":"b-39c6b324fcb45dc258f2d5ec","operation":"retain-source-identity"}],"blocks":[{"id":"b-256c7bedef66c034c7868751","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-256c7bedef66c034c7868751","content":"# CLS 08 · 密集预测、实例与集合输出","sha256":"e4f817838c06f025f95e906deb158142e9364a9ad54a95e0fa2fdf2414eeb1d4","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-opening","title":"CLS 08 · 密集预测、实例与集合输出","number":1},"sectionTitle":"CLS 08 · 密集预测、实例与集合输出","currentQuestion":"CLS 08 · 密集预测、实例与集合输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-ad4b26c04754f893d870d4c3","prerequisites":[{"id":"b-ad4b26c04754f893d870d4c3","url":"#b-ad4b26c04754f893d870d4c3","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-ad4b26c04754f893d870d4c3","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-ad4b26c04754f893d870d4c3","content":"按问题继续","sha256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-opening","title":"CLS 08 · 密集预测、实例与集合输出","number":1},"sectionTitle":"CLS 08 · 密集预测、实例与集合输出","currentQuestion":"CLS 08 · 密集预测、实例与集合输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-256c7bedef66c034c7868751","nextBlock":"b-4cb0c35f5239fbcfba4d674a","prerequisites":[{"id":"b-256c7bedef66c034c7868751","url":"#b-256c7bedef66c034c7868751","role":"section-title"}],"relatedBlocks":[]},{"id":"b-4cb0c35f5239fbcfba4d674a","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-4cb0c35f5239fbcfba4d674a","content":"像素损失、IoU口径、粗细表示、实例身份与集合输出","sha256":"70eec56f79b131d4695a1c420bc695a02ef3863727481994de8f7ed4bd673dcc","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-opening","title":"CLS 08 · 密集预测、实例与集合输出","number":1},"sectionTitle":"CLS 08 · 密集预测、实例与集合输出","currentQuestion":"CLS 08 · 密集预测、实例与集合输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ad4b26c04754f893d870d4c3","nextBlock":"b-3feef70df269f2b7ae546bef","prerequisites":[{"id":"b-256c7bedef66c034c7868751","url":"#b-256c7bedef66c034c7868751","role":"section-title"},{"id":"b-ad4b26c04754f893d870d4c3","url":"#b-ad4b26c04754f893d870d4c3","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-3feef70df269f2b7ae546bef","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-3feef70df269f2b7ae546bef","content":"可先阅读已有结果，再决定是否运行。重算时请核对各节记录的数据、参数、环境与日期。","sha256":"f22152fd55dcdb0090b5653d37145694fd19388d31e10226549464fe8b483578","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-opening","title":"CLS 08 · 密集预测、实例与集合输出","number":1},"sectionTitle":"CLS 08 · 密集预测、实例与集合输出","currentQuestion":"CLS 08 · 密集预测、实例与集合输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-4cb0c35f5239fbcfba4d674a","nextBlock":null,"prerequisites":[{"id":"b-256c7bedef66c034c7868751","url":"#b-256c7bedef66c034c7868751","role":"section-title"},{"id":"b-ad4b26c04754f893d870d4c3","url":"#b-ad4b26c04754f893d870d4c3","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7bf8697770ccfefdf868ec31","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-7bf8697770ccfefdf868ec31","content":"**学习路径：** [像素、边界与对象身份](/course/study/pixels-instances.html)","sha256":"30ab9f304366670818536cf73c98c66aa814e8e31ecdbffcf19a4d6e7e7f22de","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-7824015c8fa9007ea0c7f8e4","prerequisites":[],"relatedBlocks":[]},{"id":"b-7824015c8fa9007ea0c7f8e4","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-7824015c8fa9007ea0c7f8e4","content":"本册承接 [P36 的逐像素任务](/course/pages/l3-r2-p36.html)，解释 [LAB 11](https://deepnote.com/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/9e268706471944e3a59e6f568f9ce9a3) 为什么需要新的目标、损失和评价。先读 logits/目标/ignore，再用 8×8 算例区分前景 IoU、两类 mIoU 和像素准确率；最后看语义类别为何不能唯一给出实例身份。","sha256":"72798cbfc37e71e8b9dbbcfe3cfb8a7b0fc2d74531b251cdcb3e0c9989b9c37d","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7bf8697770ccfefdf868ec31","nextBlock":"b-e9bb6e3b3e8ff1a4f99a460d","prerequisites":[{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-e9bb6e3b3e8ff1a4f99a460d","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-e9bb6e3b3e8ff1a4f99a460d","content":"希望先看结果，可读 [RGB 读出与 FCN 的对照](/course/blocks/pixel-model-positive-result.html)；希望核一个指标，可引用 [0.667 与 0.75 的完整算例](/course/blocks/mask-iou-worked-example.html)。读完应能解释“框对了但边界仍错”和“所有杯子像素对了但实例仍未分开”这两个缺口。","sha256":"93743e4257248bf9219358b363aae08572036e7ccafb37b6917bd45eeacfbe96","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7824015c8fa9007ea0c7f8e4","nextBlock":"b-c06dd4bf7393cdb19ca92b7b","prerequisites":[{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-c06dd4bf7393cdb19ca92b7b","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-c06dd4bf7393cdb19ca92b7b","content":"坐标采样的动手入口是 [LAB 03](https://deepnote.com/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/49a2c030eafa4e1db6fc4f1a08c3810a)；多尺度和集合匹配留在本册后部选读。它们不要求在理解像素预测前先完成。","sha256":"ad5c18536ebaaed1683a9cbbc42e20e2f10a94f3ad582f81270f89de1719a98f","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e9bb6e3b3e8ff1a4f99a460d","nextBlock":"b-9f9fa2f87c514087cde4da29","prerequisites":[{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-9f9fa2f87c514087cde4da29","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-9f9fa2f87c514087cde4da29","content":"一个框可以同时包住杯身、杯把和杯把中的背景。框坐标全对，仍不代表我们知道每个像素属于谁。若两只杯子接触，语义分割可以把它们都标成“杯子”，但要分别计数和操作，就还需区分实例身份。这里不是同一个答案换显示颜色，而是输出空间和监督单位发生了变化。","sha256":"5409b0105bb5ec9dc3241c580e59c54772c6e8b9d8dbcfe5caf959343ea238a6","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c06dd4bf7393cdb19ca92b7b","nextBlock":"b-b07b8d541342c12da8f0a3f7","prerequisites":[{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-b07b8d541342c12da8f0a3f7","type":"h2","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-b07b8d541342c12da8f0a3f7","content":"## 一张类别图怎样成为可训练的输出","sha256":"77f6d53302f1be4d6d3ee57e02b21459f56e0aa208d6bbbd81accd9d8b377d53","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9f9fa2f87c514087cde4da29","nextBlock":"b-a6d93d405af42d1c53096d4c","prerequisites":[{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-a6d93d405af42d1c53096d4c","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-a6d93d405af42d1c53096d4c","content":"语义分割的输出通常为 $Z\\in\\mathbb R^{N\\times C\\times H\\times W}$，其中 $N$ 是 batch，$C$ 是类别数，$H,W$ 是空间大小。目标 $Y\\in\\{0,\\ldots,C-1\\}^{N\\times H\\times W}$ 是整数类别图。对每个有效像素做 softmax 交叉熵，再按有效像素数平均：","sha256":"3e56a076a7b9a48f3f8c0e4a1ec3518b7589369f6099f713d3ab613354f864c0","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"Z\\in\\mathbb R^{N\\times C\\times H\\times W}","definitionContext":"一张类别图怎样成为可训练的输出"},{"latex":"N","definitionContext":"一张类别图怎样成为可训练的输出"},{"latex":"C","definitionContext":"一张类别图怎样成为可训练的输出"},{"latex":"H,W","definitionContext":"一张类别图怎样成为可训练的输出"},{"latex":"Y\\in\\{0,\\ldots,C-1\\}^{N\\times H\\times W}","definitionContext":"一张类别图怎样成为可训练的输出"}],"previousBlock":"b-b07b8d541342c12da8f0a3f7","nextBlock":"b-143c0b332ba51e83ccc3baef","prerequisites":[{"id":"b-b07b8d541342c12da8f0a3f7","url":"#b-b07b8d541342c12da8f0a3f7","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-143c0b332ba51e83ccc3baef","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-143c0b332ba51e83ccc3baef","content":"$L=-\\frac{1}{|\\Omega|}\\sum_{(n,u,v)\\in\\Omega}\\log p_{n,Y_{nuv},u,v}.$","sha256":"ad790b051493beed5374493af79f9514dc8aa5e0b22af3b3dc4d840d31e23fe3","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"L=-\\frac{1}{|\\Omega|}\\sum_{(n,u,v)\\in\\Omega}\\log p_{n,Y_{nuv},u,v}.","definitionContext":"一张类别图怎样成为可训练的输出"}],"previousBlock":"b-a6d93d405af42d1c53096d4c","nextBlock":"b-65acb3f8c752f24c2f082610","prerequisites":[{"id":"b-b07b8d541342c12da8f0a3f7","url":"#b-b07b8d541342c12da8f0a3f7","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-65acb3f8c752f24c2f082610","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-65acb3f8c752f24c2f082610","content":"$\\Omega$ 排除 ignore label，例如未标注区域。ignore 不应变成一个额外背景类，也不进入评价混淆矩阵。二类任务可以用一个 foreground logit 加 sigmoid/BCE，也可以用两个 logits 加 softmax；张量形状和损失接口必须与这种选择一致。","sha256":"be0aaf3d04041ca7d223159fab6db6c9adaac7193b5a1788872d4d17ef5a188b","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"\\Omega","definitionContext":"一张类别图怎样成为可训练的输出"}],"previousBlock":"b-143c0b332ba51e83ccc3baef","nextBlock":"b-bc6e32b035fa71a395b40263","prerequisites":[{"id":"b-b07b8d541342c12da8f0a3f7","url":"#b-b07b8d541342c12da8f0a3f7","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-bc6e32b035fa71a395b40263","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-bc6e32b035fa71a395b40263","content":"FCN 用卷积保持空间网格并输出位置相关分数，再把较粗分辨率恢复到目标大小。U-Net 的编码—解码结构把深层上下文与浅层高分辨率特征通过跳连结合；这种连接提供额外可用信息，而不仅是把同一张粗图放大。[D2L §14.11 FCN](https://d2l.ai/chapter_computer-vision/fcn.html) · [U-Net 原论文](https://arxiv.org/abs/1505.04597)。图像和 mask 的裁剪、翻转、缩放必须使用同一几何变换，否则网络被要求把一个位置的像素对应到另一个位置的标签。","sha256":"a9149806afd622bccc16835a043e5f4608c2cf76dcabe0e7596a0cc5c2bf8c51","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-65acb3f8c752f24c2f082610","nextBlock":"b-8fdcd1281b281fe4293dc8d1","prerequisites":[{"id":"b-b07b8d541342c12da8f0a3f7","url":"#b-b07b8d541342c12da8f0a3f7","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-8fdcd1281b281fe4293dc8d1","type":"h2","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-8fdcd1281b281fe4293dc8d1","content":"## 上采样能增加格子，不能单凭格子恢复被删掉的区别","sha256":"241bf275da03aa8d9391b79ee6cd44b57a829f6baa38512357bc65f453f808df","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-bc6e32b035fa71a395b40263","nextBlock":"b-4fa0186098005215f085fe0c","prerequisites":[{"id":"b-b07b8d541342c12da8f0a3f7","url":"#b-b07b8d541342c12da8f0a3f7","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-4fa0186098005215f085fe0c","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-4fa0186098005215f085fe0c","content":"最近邻复制相邻值；双线性插值对邻域数值加权。给二维网格","sha256":"cf7413ba356779e875fe91f4f34c7c315f7f7b7ce42e19b44ffd1dcb30d0d58b","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-8fdcd1281b281fe4293dc8d1","nextBlock":"b-6d725477d10319e2c54bdad8","prerequisites":[{"id":"b-8fdcd1281b281fe4293dc8d1","url":"#b-8fdcd1281b281fe4293dc8d1","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-6d725477d10319e2c54bdad8","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-6d725477d10319e2c54bdad8","content":"$F=\\begin{pmatrix}0&10\\\\20&30\\end{pmatrix},$","sha256":"041babd888393a13f260ef325bcc31c4c9cd1e4f7df2491e324ddafb2b56cd23","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"F=\\begin{pmatrix}0&10\\\\20&30\\end{pmatrix},","definitionContext":"一张类别图怎样成为可训练的输出"}],"previousBlock":"b-4fa0186098005215f085fe0c","nextBlock":"b-a43c41cc687b47b717ec77c9","prerequisites":[{"id":"b-8fdcd1281b281fe4293dc8d1","url":"#b-8fdcd1281b281fe4293dc8d1","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-a43c41cc687b47b717ec77c9","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-a43c41cc687b47b717ec77c9","content":"在横坐标 $x=0.25$、纵坐标 $y=0.5$ 处，四点权重为 $0.375,0.125,0.375,0.125$，插值得到 12.5。它精确说明怎样从已有数值取连续位置，不能保证 12.5 就是原始图像那一点的真实内容。","sha256":"c315ee35160727db2c7bb7bef724e45266800f78ea5d2880c7d3101479ad3b43","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"x=0.25","definitionContext":"一张类别图怎样成为可训练的输出"},{"latex":"y=0.5","definitionContext":"一张类别图怎样成为可训练的输出"},{"latex":"0.375,0.125,0.375,0.125","definitionContext":"一张类别图怎样成为可训练的输出"}],"previousBlock":"b-6d725477d10319e2c54bdad8","nextBlock":"b-ea8636c8daedf08fdeff6409","prerequisites":[{"id":"b-8fdcd1281b281fe4293dc8d1","url":"#b-8fdcd1281b281fe4293dc8d1","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-ea8636c8daedf08fdeff6409","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-ea8636c8daedf08fdeff6409","content":"如果两个 $2\\times2$ 图案“左半亮”和“右半亮”都被平均成 0.5，任何只拿这个平均值的确定性解码器都必须给二者相同输出。增加一个“左减右”通道后才可以区分。这是 [LAB 03](https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/49a2c030eafa4e1db6fc4f1a08c3810a?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=49a2c030eafa4e1db6fc4f1a08c3810a&utm_term=get_notebook) 的表示碰撞例；它解释为什么需要保留空间信息，而非只提高显示分辨率。","sha256":"9d6d23d8e3c027ed4f0115e54b8a2a363b6e2f8a619951c7433c03f79bcd0992","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"2\\times2","definitionContext":"一张类别图怎样成为可训练的输出"}],"previousBlock":"b-a43c41cc687b47b717ec77c9","nextBlock":"b-d7e7e9870d55b802a8b461d1","prerequisites":[{"id":"b-8fdcd1281b281fe4293dc8d1","url":"#b-8fdcd1281b281fe4293dc8d1","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-d7e7e9870d55b802a8b461d1","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-d7e7e9870d55b802a8b461d1","content":"类别 mask 本身是离散 ID，通常用最近邻几何采样。对类别编号 1 和 3 做双线性插值，得到 2 并不意味着中间真的存在第 2 类。模型的连续 logits 或概率图则可以按适当约定插值，再 argmax/阈值。插值坐标约定也要固定，例如是否对齐端点；同样叫 bilinear 的两个设置可能在边界位置不同。","sha256":"ae393bf00717f2b8ab32d920a04af01319f2e21853fbca2d696f4bb48f770610","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ea8636c8daedf08fdeff6409","nextBlock":"b-47a7abca10abbc6fc0131e80","prerequisites":[{"id":"b-8fdcd1281b281fe4293dc8d1","url":"#b-8fdcd1281b281fe4293dc8d1","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-47a7abca10abbc6fc0131e80","type":"h2","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-47a7abca10abbc6fc0131e80","content":"## 一个 8×8 反例怎样澄清四种指标","sha256":"35d7aa5a6598e52321e9309a0916b06dc03cd25c5ad4df6383272b464361299a","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d7e7e9870d55b802a8b461d1","nextBlock":"b-58c643129233b5bf3fcdf363","prerequisites":[{"id":"b-8fdcd1281b281fe4293dc8d1","url":"#b-8fdcd1281b281fe4293dc8d1","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-58c643129233b5bf3fcdf363","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-58c643129233b5bf3fcdf363","content":"设真值前景有 16 像素，从真值 mask 平均降采样再最近邻恢复并阈值化后，预测前景有 24 像素，其中真阳性 16、假阳性 8、假阴性 0、真阴性 40。这是采样机制例，预测过程直接使用了真值 mask，不能当作看图模型的能力证据。","sha256":"cdf5e1ee80a89ea700fef3561ecdfa2d3713ae806370d2924e5897e2a5078529","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-47a7abca10abbc6fc0131e80","nextBlock":"b-979030e464e9fa7cadcd3fa4","prerequisites":[{"id":"b-47a7abca10abbc6fc0131e80","url":"#b-47a7abca10abbc6fc0131e80","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-979030e464e9fa7cadcd3fa4","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-979030e464e9fa7cadcd3fa4","content":"![真值掩码、降采样恢复和错误区域](/course/repair/assets/cls09-mask-metrics.svg)","sha256":"d8c17ff28b5b0d5b1671238ddcef7a1c477eeded2b48b525dadaddb86200d640","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-58c643129233b5bf3fcdf363","nextBlock":"b-d367e0432b7c153aa011785b","prerequisites":[{"id":"b-47a7abca10abbc6fc0131e80","url":"#b-47a7abca10abbc6fc0131e80","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-d367e0432b7c153aa011785b","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-d367e0432b7c153aa011785b","content":"三张网格由同一组 8×8 数组绘制。目标矩形为行 2–5、列 3–6（从 0 计数）；阈值化最近邻恢复多出了 8 个前景像素。","sha256":"13df4d569b221f8bc60c5de960a783d1972c800851f3e5ebd251ca5ca31bb27c","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-979030e464e9fa7cadcd3fa4","nextBlock":"b-27cdb17f336635825aa19e6f","prerequisites":[{"id":"b-47a7abca10abbc6fc0131e80","url":"#b-47a7abca10abbc6fc0131e80","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-27cdb17f336635825aa19e6f","type":"table","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-27cdb17f336635825aa19e6f","content":"| 指标 | 计算 | 本例结果 |\n| --- | --- | --- |\n| 前景 IoU | $16/(16+8+0)$ | $2/3$ |\n| 背景 IoU | $40/(40+0+8)$ | $5/6$ |\n| 两类 mIoU，含背景 | $(2/3+5/6)/2$ | 0.75 |\n| 像素准确率 | $(16+40)/64$ | 0.875 |","sha256":"d12b72281bce9d27865a10e904405cca1c2c915b9910c63934ad6f7d4439e49a","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"16/(16+8+0)","definitionContext":"一张类别图怎样成为可训练的输出"},{"latex":"2/3","definitionContext":"一张类别图怎样成为可训练的输出"},{"latex":"40/(40+0+8)","definitionContext":"一张类别图怎样成为可训练的输出"},{"latex":"5/6","definitionContext":"一张类别图怎样成为可训练的输出"},{"latex":"(2/3+5/6)/2","definitionContext":"一张类别图怎样成为可训练的输出"},{"latex":"(16+40)/64","definitionContext":"一张类别图怎样成为可训练的输出"}],"table":{"headers":["指标","计算","本例结果"],"rows":[["指标","计算","本例结果"],["前景 IoU","16/(16+8+0)16/(16+8+0)16/(16+8+0)","2/32/32/3"],["背景 IoU","40/(40+0+8)40/(40+0+8)40/(40+0+8)","5/65/65/6"],["两类 mIoU，含背景","(2/3+5/6)/2(2/3+5/6)/2(2/3+5/6)/2","0.75"],["像素准确率","(16+40)/64(16+40)/64(16+40)/64","0.875"]]},"previousBlock":"b-d367e0432b7c153aa011785b","nextBlock":"b-6a5d378fb1f084605094a70a","prerequisites":[{"id":"b-47a7abca10abbc6fc0131e80","url":"#b-47a7abca10abbc6fc0131e80","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-6a5d378fb1f084605094a70a","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-6a5d378fb1f084605094a70a","content":"“mIoU=0.667”在这个例子里错在名称：0.667 是单类前景 IoU。若报告不含背景的类均值，应写清类别集合；若某类真值和预测都为空，要说明是忽略该类还是采用约定值。背景占比很高时，全预测背景也可能有高像素准确率，所以要看逐类混淆和边界错误，而不只看一个总数。","sha256":"4c2978607bb7b437fdad1528eb065e99e5968fdc50592f9201ed43dcf3ded1a3","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-27cdb17f336635825aa19e6f","nextBlock":"b-b0d1ded46e6bebb316c4efe2","prerequisites":[{"id":"b-47a7abca10abbc6fc0131e80","url":"#b-47a7abca10abbc6fc0131e80","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-b0d1ded46e6bebb316c4efe2","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-b0d1ded46e6bebb316c4efe2","content":"真正视觉模型必须从图像推断 mask。[LAB 11](https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/9e268706471944e3a59e6f568f9ce9a3?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=9e268706471944e3a59e6f568f9ce9a3&utm_term=get_notebook) 先保留上述采样例及最近邻/双线性比较，再进入小型像素模型。两段实验分别检验采样损失和学习能力，结论不互相替代。","sha256":"939045e192e80393bc29ee5638c03a5341285cb4e18440f27403c32464cc6f54","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6a5d378fb1f084605094a70a","nextBlock":"b-ec127c6207c04d9876a76d11","prerequisites":[{"id":"b-47a7abca10abbc6fc0131e80","url":"#b-47a7abca10abbc6fc0131e80","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-ec127c6207c04d9876a76d11","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-ec127c6207c04d9876a76d11","content":"主实验在 2026-09-22 本地 CPU 完成训练，384 张训练图、128 张保留图，输入 48×48 RGB。三类为背景、红矩形、蓝椭圆，目标形状为 `[B,48,48]`，logits 为 `[B,3,48,48]`，ignore label 255 在损失与指标中均被排除。20,590 参数的小型 FCN 将 12×12 粗 logits 双线性上采样并融合精细跳连，训练 20 epoch；另训练只有 12 个参数的逐像素 RGB 线性读出作基线。","sha256":"715899b38a78b3bf4f5526299aecfd4f7c5178ff28d2193adb58e1166be1ad9d","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b0d1ded46e6bebb316c4efe2","nextBlock":"b-7987bf989df1a4f31935446a","prerequisites":[{"id":"b-47a7abca10abbc6fc0131e80","url":"#b-47a7abca10abbc6fc0131e80","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7987bf989df1a4f31935446a","type":"table","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-7987bf989df1a4f31935446a","content":"| 保留图像条件 | 小型 FCN 三类 mIoU | RGB 逐像素线性读出 mIoU |\n| --- | --- | --- |\n| 原始合成图 | 0.99041 | 1.00000 |\n| 对象尺度变为 0.45 倍 | 0.94396 | 1.00000 |\n| RGB 图作 5×5 模糊 | 0.92397 | 0.84996 |","sha256":"506f6e140199a9480864e32072bf7dc411386e1f8072f90aaf04d4a7616bf87e","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"table":{"headers":["保留图像条件","小型 FCN 三类 mIoU","RGB 逐像素线性读出 mIoU"],"rows":[["保留图像条件","小型 FCN 三类 mIoU","RGB 逐像素线性读出 mIoU"],["原始合成图","0.99041","1.00000"],["对象尺度变为 0.45 倍","0.94396","1.00000"],["RGB 图作 5×5 模糊","0.92397","0.84996"]]},"previousBlock":"b-ec127c6207c04d9876a76d11","nextBlock":"b-a44a66cfeed574e75874bced","prerequisites":[{"id":"b-47a7abca10abbc6fc0131e80","url":"#b-47a7abca10abbc6fc0131e80","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-a44a66cfeed574e75874bced","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-a44a66cfeed574e75874bced","content":"RGB 基线满分说明原数据颜色已直接泄露类别，不能据 FCN 高分宣称它学会了复杂形状或需要深层上下文。模糊后边缘颜色混合，FCN 在这个干预下更稳健；这才是当前对照支持的具体比较。原始条件下 FCN 的逐类 IoU 为背景 0.99861、矩形 0.99833、椭圆 0.97429，较细曲线边界更难。","sha256":"596f845a43d7a997861347a6bbdfe900ef177b21daf6e5f5bc7a191ee6b9e7f8","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7987bf989df1a4f31935446a","nextBlock":"b-dfb5a10e7e678b16d9e2eb51","prerequisites":[{"id":"b-47a7abca10abbc6fc0131e80","url":"#b-47a7abca10abbc6fc0131e80","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-dfb5a10e7e678b16d9e2eb51","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-dfb5a10e7e678b16d9e2eb51","content":"![真实图像、mask、FCN输出与错误](/assets/display/f2028b85a049b53efbcbbb7a.webp)","sha256":"f56f575033bcab034622fd5bcf423ca70f6f99a0671bfa0c2557bc4ade38d310","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-a44a66cfeed574e75874bced","nextBlock":"b-1fa6ef11953df3de99116beb","prerequisites":[{"id":"b-47a7abca10abbc6fc0131e80","url":"#b-47a7abca10abbc6fc0131e80","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-1fa6ef11953df3de99116beb","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-1fa6ef11953df3de99116beb","content":"![小对象条件下的错误分布](/assets/display/a3b94d7df0f7bd4472d4f87d.webp)","sha256":"6e59fe129adbd7f631cd6517b8715f8e88361fc81316790d51a971ff6a7ac9d1","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-dfb5a10e7e678b16d9e2eb51","nextBlock":"b-5c19a3719eb906055e61116c","prerequisites":[{"id":"b-47a7abca10abbc6fc0131e80","url":"#b-47a7abca10abbc6fc0131e80","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-5c19a3719eb906055e61116c","type":"p","scope":"block-6ca34c15abaf4aeb9900af72759ebc28","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-5c19a3719eb906055e61116c","content":"保持已训练参数，测试时仅把粗 logits 的双线性插值改为最近邻，mIoU 降到 0.91180。这说明当前网络对训练时的采样约定敏感，不能把它读作“最近邻永远更差”，因为网络没有在两种上采样设置下分别重训。前面的已知 mask 热身中，双线性分支前景 IoU 仍为 2/3，背景 IoU 为 0.88462、两类 mIoU 为 0.77564；指标差异与模型泛化仍须分开。","sha256":"239d41a1854eaf46c28660b8f1fafbd9e2dd5243ba3e50949f8b9548fb79b67f","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-6ca34c15abaf4aeb9900af72759ebc28","title":"一张类别图怎样成为可训练的输出","number":2},"sectionTitle":"一张类别图怎样成为可训练的输出","currentQuestion":"一张类别图怎样成为可训练的输出","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-1fa6ef11953df3de99116beb","nextBlock":null,"prerequisites":[{"id":"b-47a7abca10abbc6fc0131e80","url":"#b-47a7abca10abbc6fc0131e80","role":"section-title"},{"id":"b-7bf8697770ccfefdf868ec31","url":"#b-7bf8697770ccfefdf868ec31","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-88083cc9f29886ff67d15c35","type":"h2","scope":"block-8413a612fc314b05806c00e9cf00fb7a","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-88083cc9f29886ff67d15c35","content":"## 从类别像素走向逐个对象","sha256":"e4fe0386a7ec4009e0b726ed0380c146b413771e54cc9b67fa42f3c17ddc9e69","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","number":3},"sectionTitle":"从类别像素走向逐个对象","currentQuestion":"从类别像素走向逐个对象","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-e4182d7601f80e13d00c0c32","prerequisites":[{"id":"b-e4182d7601f80e13d00c0c32","url":"#b-e4182d7601f80e13d00c0c32","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-e4182d7601f80e13d00c0c32","type":"p","scope":"block-8413a612fc314b05806c00e9cf00fb7a","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-e4182d7601f80e13d00c0c32","content":"把像素标为杯子类别，只声明“这里属于杯子集合”。实例标注还声明“这里属于杯一，那里属于杯二”。两只杯子即使同色、接触甚至相互遮挡，实例任务仍需给出两份对象身份。","sha256":"62cd3c3e7472a4976a8c101608c68d04df6eb55e6ed8b1cdfe8b808e43846633","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","number":3},"sectionTitle":"从类别像素走向逐个对象","currentQuestion":"从类别像素走向逐个对象","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-88083cc9f29886ff67d15c35","nextBlock":"b-59d22d3f12d4a46fa39d9236","prerequisites":[{"id":"b-88083cc9f29886ff67d15c35","url":"#b-88083cc9f29886ff67d15c35","role":"section-title"}],"relatedBlocks":[]},{"id":"b-59d22d3f12d4a46fa39d9236","type":"p","scope":"block-8413a612fc314b05806c00e9cf00fb7a","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-59d22d3f12d4a46fa39d9236","content":"考虑一行标签 `[1,1,2,2]`，其中 1、2 是两只杯子的实例号。将它们都映射到类别杯子=1，就得到 `[1,1,1,1]`。另一个实例划分 `[1,2,2,2]` 经同样映射也得到同一类别行。因此类别图不能唯一反推出实例边界。","sha256":"b77c69f411f0cec66906efa93953096b72fee0517dd04ceb2eb4f2c14fca2c6c","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","number":3},"sectionTitle":"从类别像素走向逐个对象","currentQuestion":"从类别像素走向逐个对象","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e4182d7601f80e13d00c0c32","nextBlock":"b-e791c92f48d99504eec4e562","prerequisites":[{"id":"b-88083cc9f29886ff67d15c35","url":"#b-88083cc9f29886ff67d15c35","role":"section-title"},{"id":"b-e4182d7601f80e13d00c0c32","url":"#b-e4182d7601f80e13d00c0c32","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-e791c92f48d99504eec4e562","type":"p","scope":"block-8413a612fc314b05806c00e9cf00fb7a","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-e791c92f48d99504eec4e562","content":"连通域可以在“每个对象互不接触”的特定条件下帮忙计数，但一旦两只杯子相接，一个连通块可含两个对象；反过来同一被遮挡的对象还可能产生不连通的可见部分。对象身份不是连通性规则自动保证的属性。","sha256":"df469b42a008cfb7ecd8a88fd7f7836c4bc0779d31028b4938705dc288908ff6","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","number":3},"sectionTitle":"从类别像素走向逐个对象","currentQuestion":"从类别像素走向逐个对象","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-59d22d3f12d4a46fa39d9236","nextBlock":"b-763d584a8643606e94939c2b","prerequisites":[{"id":"b-88083cc9f29886ff67d15c35","url":"#b-88083cc9f29886ff67d15c35","role":"section-title"},{"id":"b-e4182d7601f80e13d00c0c32","url":"#b-e4182d7601f80e13d00c0c32","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-763d584a8643606e94939c2b","type":"p","scope":"block-8413a612fc314b05806c00e9cf00fb7a","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-763d584a8643606e94939c2b","content":"区域路线给每个候选一份身份，像素路线给精细轮廓。把它们结合，就可以在“这个候选对象”之内预测局部 mask。这是进入 Mask R-CNN 的具体需求，而不是为了再增加一种模型名称。","sha256":"ea717067584e344418a930c49862e5bea35e54c3e75453a5dc35674aa7112bb5","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","number":3},"sectionTitle":"从类别像素走向逐个对象","currentQuestion":"从类别像素走向逐个对象","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e791c92f48d99504eec4e562","nextBlock":"b-9b2fdf37cd22b7d4e05509d9","prerequisites":[{"id":"b-88083cc9f29886ff67d15c35","url":"#b-88083cc9f29886ff67d15c35","role":"section-title"},{"id":"b-e4182d7601f80e13d00c0c32","url":"#b-e4182d7601f80e13d00c0c32","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-9b2fdf37cd22b7d4e05509d9","type":"p","scope":"block-8413a612fc314b05806c00e9cf00fb7a","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-9b2fdf37cd22b7d4e05509d9","content":"[Mask R-CNN 的三项输出与局部目标](/course/blocks/mask-branch-target.html) · [同一杯子换RoI后，目标如何改变](/course/blocks/local-mask-coordinate-target.html) · [P40](/course/pages/l3-r2-p40.html)—[P43](/course/pages/l3-r2-p43.html)。","sha256":"fdde3520bab89e32b78dca4f14b82d41aa943c01e54d2b8a789f56ae0adbb053","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","number":3},"sectionTitle":"从类别像素走向逐个对象","currentQuestion":"从类别像素走向逐个对象","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-763d584a8643606e94939c2b","nextBlock":"b-1467ea4c04cf6b2f6f3bad13","prerequisites":[{"id":"b-88083cc9f29886ff67d15c35","url":"#b-88083cc9f29886ff67d15c35","role":"section-title"},{"id":"b-e4182d7601f80e13d00c0c32","url":"#b-e4182d7601f80e13d00c0c32","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-1467ea4c04cf6b2f6f3bad13","type":"h4","scope":"block-8413a612fc314b05806c00e9cf00fb7a","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-1467ea4c04cf6b2f6f3bad13","content":"#### 选读：不同尺度与无序集合，分别改变哪一部分？","sha256":"86b78c60de3cb89b29e5e15feabf0eec6ebc9dc9cdb49c1a74515ab154cd005f","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","number":3},"sectionTitle":"从类别像素走向逐个对象","currentQuestion":"从类别像素走向逐个对象","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9b2fdf37cd22b7d4e05509d9","nextBlock":"b-5bdd749d5bc44be721f1010a","prerequisites":[{"id":"b-88083cc9f29886ff67d15c35","url":"#b-88083cc9f29886ff67d15c35","role":"section-title"},{"id":"b-e4182d7601f80e13d00c0c32","url":"#b-e4182d7601f80e13d00c0c32","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-5bdd749d5bc44be721f1010a","type":"p","scope":"block-8413a612fc314b05806c00e9cf00fb7a","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-5bdd749d5bc44be721f1010a","content":"把“有跳连”当成一个共同名字，容易遮住任务差异。FCN 的跳接把较细层的空间线索与粗语义预测结合，服务于逐像素输出；U-Net 的编码—解码路径将高分辨率特征接到相应解码层；FPN 自顶向下融合与横向连接，形成可供不同尺度对象使用的特征层级。","sha256":"e9fee9ad6cac4094a8f46cb3106e8782d3fee4eecd1c1f2207ca52af17c2dec0","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","number":3},"sectionTitle":"从类别像素走向逐个对象","currentQuestion":"从类别像素走向逐个对象","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-1467ea4c04cf6b2f6f3bad13","nextBlock":"b-84e1878787815e7595dda972","prerequisites":[{"id":"b-1467ea4c04cf6b2f6f3bad13","url":"#b-1467ea4c04cf6b2f6f3bad13","role":"section-title"},{"id":"b-e4182d7601f80e13d00c0c32","url":"#b-e4182d7601f80e13d00c0c32","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-84e1878787815e7595dda972","type":"table","scope":"block-8413a612fc314b05806c00e9cf00fb7a","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-84e1878787815e7595dda972","content":"| 观察点 | 应在计算图里找什么 |\n| --- | --- |\n| 较细信息从哪里来 | 原图或编码器哪个中间层，而非只看上采样箭头 |\n| 怎么对齐 | 两条支路的步长、尺寸、坐标约定 |\n| 怎样组合 | 相加或拼接，组合后是否再有可学习预测 |\n| 服务哪个目标 | 单张类别图、多尺度框、局部实例 mask |","sha256":"e2deae2a61518dc33f64cf5fcf53f9feb4901c2f5cf511ac72edb945dd9ab022","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","number":3},"sectionTitle":"从类别像素走向逐个对象","currentQuestion":"从类别像素走向逐个对象","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"table":{"headers":["观察点","应在计算图里找什么"],"rows":[["观察点","应在计算图里找什么"],["较细信息从哪里来","原图或编码器哪个中间层，而非只看上采样箭头"],["怎么对齐","两条支路的步长、尺寸、坐标约定"],["怎样组合","相加或拼接，组合后是否再有可学习预测"],["服务哪个目标","单张类别图、多尺度框、局部实例 mask"]]},"previousBlock":"b-5bdd749d5bc44be721f1010a","nextBlock":"b-6eabbd70e912c0ae34c10cb3","prerequisites":[{"id":"b-1467ea4c04cf6b2f6f3bad13","url":"#b-1467ea4c04cf6b2f6f3bad13","role":"section-title"},{"id":"b-e4182d7601f80e13d00c0c32","url":"#b-e4182d7601f80e13d00c0c32","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-6eabbd70e912c0ae34c10cb3","type":"p","scope":"block-8413a612fc314b05806c00e9cf00fb7a","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-6eabbd70e912c0ae34c10cb3","content":"如果原图中 4×4 像素的目标进入 stride16 网格，其连续跨度只有 1/4×1/4 格；这使位置读取困难，但不能仅据小于一格断言信息必全丢，通道仍可能记录响应。更细输入或跨层特征带来新的观测机会，是否改善要在指定任务上检查。","sha256":"b2da5a7d8caaa632c7ae9f9f6ce0a06fe464eecb40e6472dfe0da761208eb634","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","number":3},"sectionTitle":"从类别像素走向逐个对象","currentQuestion":"从类别像素走向逐个对象","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-84e1878787815e7595dda972","nextBlock":"b-112423ff84ed57e9e701e5b3","prerequisites":[{"id":"b-1467ea4c04cf6b2f6f3bad13","url":"#b-1467ea4c04cf6b2f6f3bad13","role":"section-title"},{"id":"b-e4182d7601f80e13d00c0c32","url":"#b-e4182d7601f80e13d00c0c32","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-112423ff84ed57e9e701e5b3","type":"p","scope":"block-8413a612fc314b05806c00e9cf00fb7a","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-112423ff84ed57e9e701e5b3","content":"保留这份比较，是为学生读到结构名时有判断坐标；本轮主讲不再把它拆为几个平行论文段落。","sha256":"4f70352dd5540e712df3eec48a1d8a256a043e13993dd8de8b43b7b9f446b8b8","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","number":3},"sectionTitle":"从类别像素走向逐个对象","currentQuestion":"从类别像素走向逐个对象","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6eabbd70e912c0ae34c10cb3","nextBlock":"b-fcfea371257d89215d96ac41","prerequisites":[{"id":"b-1467ea4c04cf6b2f6f3bad13","url":"#b-1467ea4c04cf6b2f6f3bad13","role":"section-title"},{"id":"b-e4182d7601f80e13d00c0c32","url":"#b-e4182d7601f80e13d00c0c32","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-fcfea371257d89215d96ac41","type":"p","scope":"block-8413a612fc314b05806c00e9cf00fb7a","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-fcfea371257d89215d96ac41","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":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","number":3},"sectionTitle":"从类别像素走向逐个对象","currentQuestion":"从类别像素走向逐个对象","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"(1-p_t)^\\gamma","definitionContext":"从类别像素走向逐个对象"}],"previousBlock":"b-112423ff84ed57e9e701e5b3","nextBlock":"b-7166a465c2c5e0f2b4734595","prerequisites":[{"id":"b-1467ea4c04cf6b2f6f3bad13","url":"#b-1467ea4c04cf6b2f6f3bad13","role":"section-title"},{"id":"b-e4182d7601f80e13d00c0c32","url":"#b-e4182d7601f80e13d00c0c32","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7166a465c2c5e0f2b4734595","type":"p","scope":"block-8413a612fc314b05806c00e9cf00fb7a","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-7166a465c2c5e0f2b4734595","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":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","number":3},"sectionTitle":"从类别像素走向逐个对象","currentQuestion":"从类别像素走向逐个对象","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-fcfea371257d89215d96ac41","nextBlock":"b-acd83acacdc33fad32e43c4a","prerequisites":[{"id":"b-1467ea4c04cf6b2f6f3bad13","url":"#b-1467ea4c04cf6b2f6f3bad13","role":"section-title"},{"id":"b-e4182d7601f80e13d00c0c32","url":"#b-e4182d7601f80e13d00c0c32","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-acd83acacdc33fad32e43c4a","type":"p","scope":"block-8413a612fc314b05806c00e9cf00fb7a","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-acd83acacdc33fad32e43c4a","content":"LAB 20 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":"6ccfb455e3b10fd39f061e5a79711a72b3121ff01414683996e829721f063b92","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","number":3},"sectionTitle":"从类别像素走向逐个对象","currentQuestion":"从类别像素走向逐个对象","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7166a465c2c5e0f2b4734595","nextBlock":"b-183259a5cc28aa462c2bff45","prerequisites":[{"id":"b-1467ea4c04cf6b2f6f3bad13","url":"#b-1467ea4c04cf6b2f6f3bad13","role":"section-title"},{"id":"b-e4182d7601f80e13d00c0c32","url":"#b-e4182d7601f80e13d00c0c32","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-183259a5cc28aa462c2bff45","type":"p","scope":"block-8413a612fc314b05806c00e9cf00fb7a","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-183259a5cc28aa462c2bff45","content":"[LAB 20 尺度与集合短算例](https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/9315996660f54e92847770355b648f9e?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=9315996660f54e92847770355b648f9e&utm_term=get_notebook) 保留实际成本构造；不需要为了理解像素目标先学完 DETR。完整注意力结构留在第五讲，不在这里展开。","sha256":"a27c6258b1428ede911a5cd9e234dac0995bdd9caa159cef345082b9ebf07a51","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-8413a612fc314b05806c00e9cf00fb7a","title":"从类别像素走向逐个对象","number":3},"sectionTitle":"从类别像素走向逐个对象","currentQuestion":"从类别像素走向逐个对象","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-acd83acacdc33fad32e43c4a","nextBlock":null,"prerequisites":[{"id":"b-1467ea4c04cf6b2f6f3bad13","url":"#b-1467ea4c04cf6b2f6f3bad13","role":"section-title"},{"id":"b-e4182d7601f80e13d00c0c32","url":"#b-e4182d7601f80e13d00c0c32","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-39c6b324fcb45dc258f2d5ec","type":"h2","scope":"block-59e4b23da62d4b709cc87f278811ec2f","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-39c6b324fcb45dc258f2d5ec","content":"## 对齐的正面实证","sha256":"44c753d27f6a83b57ddfb3587ff00d4a897cf8a6fa52e22d06bc9714f5b85014","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-59e4b23da62d4b709cc87f278811ec2f","title":"对齐的正面实证","number":4},"sectionTitle":"对齐的正面实证","currentQuestion":"对齐的正面实证","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-9f136b9fb2d77584285a5379","prerequisites":[{"id":"b-9f136b9fb2d77584285a5379","url":"#b-9f136b9fb2d77584285a5379","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-9f136b9fb2d77584285a5379","type":"p","scope":"block-59e4b23da62d4b709cc87f278811ec2f","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-9f136b9fb2d77584285a5379","content":"Mask R-CNN v3 的 Table 2 使用 COCO `trainval35k` 训练、`minival` 测试；表中未另说明时报告 mask AP。以下固定 ResNet-50-C4、stride 16，并选同为 max 聚合的行：","sha256":"3bb8d279d8595e2246936d624e3b4edea6ceca91fece02cd289fada770a219f4","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-59e4b23da62d4b709cc87f278811ec2f","title":"对齐的正面实证","number":4},"sectionTitle":"对齐的正面实证","currentQuestion":"对齐的正面实证","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-39c6b324fcb45dc258f2d5ec","nextBlock":"b-1aee8deba52655a72cd8d405","prerequisites":[{"id":"b-39c6b324fcb45dc258f2d5ec","url":"#b-39c6b324fcb45dc258f2d5ec","role":"section-title"}],"relatedBlocks":[]},{"id":"b-1aee8deba52655a72cd8d405","type":"table","scope":"block-59e4b23da62d4b709cc87f278811ec2f","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-1aee8deba52655a72cd8d405","content":"| 区域读取 | 是否保留对齐 | 是否双线性 | mask AP | mask AP75 |\n| --- | --- | --- | --- | --- |\n| RoIPool | 否 | 否 | 26.9 | 26.4 |\n| RoIWarp | 否 | 是 | 27.2 | 27.1 |\n| RoIAlign | 是 | 是 | 30.2 | 31.8 |","sha256":"f59df6ce113db0faf2717e5b8440d85ef1f069b2e8b327b430459828cfd41119","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-59e4b23da62d4b709cc87f278811ec2f","title":"对齐的正面实证","number":4},"sectionTitle":"对齐的正面实证","currentQuestion":"对齐的正面实证","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"table":{"headers":["区域读取","是否保留对齐","是否双线性","mask AP","mask AP75"],"rows":[["区域读取","是否保留对齐","是否双线性","mask AP","mask AP75"],["RoIPool","否","否","26.9","26.4"],["RoIWarp","否","是","27.2","27.1"],["RoIAlign","是","是","30.2","31.8"]]},"previousBlock":"b-9f136b9fb2d77584285a5379","nextBlock":"b-3546af0154b89215a19aa729","prerequisites":[{"id":"b-39c6b324fcb45dc258f2d5ec","url":"#b-39c6b324fcb45dc258f2d5ec","role":"section-title"},{"id":"b-9f136b9fb2d77584285a5379","url":"#b-9f136b9fb2d77584285a5379","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-3546af0154b89215a19aa729","type":"p","scope":"block-59e4b23da62d4b709cc87f278811ec2f","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-3546af0154b89215a19aa729","content":"RoIAlign 相比 RoIPool 的 AP 增加 3.3 点，AP75 增加 5.4 点。RoIWarp 也插值却没有同样改善，支持“保持坐标对应”这一机制判断。平均聚合的 RoIAlign 行是 AP30.3/AP75 31.5，引用时不能把不同聚合行混在一个对照里。Table 2(d) 换成 C5、stride32，mask AP 从 23.6 到 30.9；这是另一组设置。","sha256":"4cdba8cde1b8262a042dd6aa8cda6abe7fd5a4e8cb6f224f5bb95e0f3611cd8d","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-59e4b23da62d4b709cc87f278811ec2f","title":"对齐的正面实证","number":4},"sectionTitle":"对齐的正面实证","currentQuestion":"对齐的正面实证","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-1aee8deba52655a72cd8d405","nextBlock":"b-926032e76429a0df5a954b76","prerequisites":[{"id":"b-39c6b324fcb45dc258f2d5ec","url":"#b-39c6b324fcb45dc258f2d5ec","role":"section-title"},{"id":"b-9f136b9fb2d77584285a5379","url":"#b-9f136b9fb2d77584285a5379","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-926032e76429a0df5a954b76","type":"p","scope":"block-59e4b23da62d4b709cc87f278811ec2f","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-926032e76429a0df5a954b76","content":"这些是训练完整模型的论文结果，不是本课程重跑。它们说明读取位置值得修复，尤其严格定位受益；P44 则把输入 F 固定成完全相同，讨论任何确定性读出都无法创造差别。方法有效与构造例中的信息限制，可以同时成立。","sha256":"a75c59c916a80252869b500277c7120d64fcb5056b16b7893076a6f49e7c3632","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-59e4b23da62d4b709cc87f278811ec2f","title":"对齐的正面实证","number":4},"sectionTitle":"对齐的正面实证","currentQuestion":"对齐的正面实证","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3546af0154b89215a19aa729","nextBlock":"b-a1af8ade7f5836656416e1e2","prerequisites":[{"id":"b-39c6b324fcb45dc258f2d5ec","url":"#b-39c6b324fcb45dc258f2d5ec","role":"section-title"},{"id":"b-9f136b9fb2d77584285a5379","url":"#b-9f136b9fb2d77584285a5379","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-a1af8ade7f5836656416e1e2","type":"p","scope":"block-59e4b23da62d4b709cc87f278811ec2f","url":"https://codingai-lec03.pages.dev/course/cls09-密集预测实例与集合输出.html#b-a1af8ade7f5836656416e1e2","content":"[Mask R-CNN v3 Table 2](https://arxiv.org/html/1703.06870v3#S4.SS2) · [对齐与信息碰撞的模拟追问](/course/blocks/roi-information-dialogue.html)","sha256":"1362761b98311ffab60c114fa4dc5d8cdf666b37458ced4e4b9d4483229b8e7e","chapter":{"id":"course","title":"CLS 08 · 密集预测、实例与集合输出"},"unit":{"id":"cls09-密集预测实例与集合输出","title":"CLS 08 · 密集预测、实例与集合输出"},"page":{"id":"reading-block-59e4b23da62d4b709cc87f278811ec2f","title":"对齐的正面实证","number":4},"sectionTitle":"对齐的正面实证","currentQuestion":"对齐的正面实证","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-926032e76429a0df5a954b76","nextBlock":null,"prerequisites":[{"id":"b-39c6b324fcb45dc258f2d5ec","url":"#b-39c6b324fcb45dc258f2d5ec","role":"section-title"},{"id":"b-9f136b9fb2d77584285a5379","url":"#b-9f136b9fb2d77584285a5379","role":"problem-context"}],"relatedBlocks":[]}]}