{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"FCN：另一条路线，直接为像素作预测","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html","markdown":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.md","pages":[{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","unit":"怎样直接学习一张像素类别地图？","unitId":"pixel-prediction-objective","chapter":"II","number":1},{"id":"pixel-supervision-contract","title":"让 logits、标签和有效像素一一对应","unit":"怎样直接学习一张像素类别地图？","unitId":"pixel-prediction-objective","chapter":"II","number":2},{"id":"pixel-lab11-baseline-evidence","title":"进入 LAB11：保留简单基线再看条件变化","unit":"哪些信息真的改善了像素预测？","unitId":"pixel-detail-and-baselines","chapter":"II","number":3}],"aliases":[{"oldId":"explanation","targetId":"pixel-fcn-route","operation":"split-explanation-range"},{"oldId":"b-fccbf303d8f895e6e5afc7cf","targetId":"b-fccbf303d8f895e6e5afc7cf","page":"pixel-fcn-route","operation":"move","beforeSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","afterSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","reason":"Same source body in authored semantic page"},{"oldId":"b-53c3821247812cc98407a346","targetId":"b-53c3821247812cc98407a346","page":"pixel-fcn-route","operation":"move","beforeSHA256":"698f36893197b1d6e08b0bf5474b9348b819281814db1411c57d710d12f43df6","afterSHA256":"698f36893197b1d6e08b0bf5474b9348b819281814db1411c57d710d12f43df6","reason":"Same source body in authored semantic page"},{"oldId":"b-cd5deee9ce543a51a693e4b7","targetId":"b-cd5deee9ce543a51a693e4b7","page":"pixel-fcn-route","operation":"move","beforeSHA256":"183d41f9f2dd4315d6651bd668ce69b215479fec798b3e287bba7d446b7f4a31","afterSHA256":"183d41f9f2dd4315d6651bd668ce69b215479fec798b3e287bba7d446b7f4a31","reason":"Same source body in authored semantic page"},{"oldId":"b-b74111f2f662cd0dcfedc6a2","targetId":"b-b74111f2f662cd0dcfedc6a2","page":"pixel-fcn-route","operation":"rewrite","beforeSHA256":"c06f1fd7f80988c85778612fa79da525232372508c0b7bbf3c535495d90c7ed7","afterSHA256":"fafc5dc0746074c8919d638becd9179587704ce91ea6431b2795bd9bae0eb43f","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-662c1d2a0499d3ca223cf2cf","targetId":"b-662c1d2a0499d3ca223cf2cf","page":"pixel-fcn-route","operation":"move","beforeSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","afterSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","reason":"Same source body in authored semantic page"},{"oldId":"b-1c949701f78f7f0e33c9873e","targetId":"b-1c949701f78f7f0e33c9873e","page":"pixel-fcn-route","operation":"move","beforeSHA256":"cd13f386087a944b59dbd751e7acf3e67a0cea054518c1212856523a6b549dc2","afterSHA256":"cd13f386087a944b59dbd751e7acf3e67a0cea054518c1212856523a6b549dc2","reason":"Same source body in authored semantic page"},{"oldId":"b-283c4597a38d3776bae3455f","targetId":"b-283c4597a38d3776bae3455f","page":"pixel-fcn-route","operation":"move","beforeSHA256":"689bfff2fcfc0e090d7c51735f916750b81432aa009d448f5600b865c9742259","afterSHA256":"689bfff2fcfc0e090d7c51735f916750b81432aa009d448f5600b865c9742259","reason":"Same source body in authored semantic page"},{"oldId":"b-5354e2cc50e944a5cfeaef6f","targetId":"b-5354e2cc50e944a5cfeaef6f","page":"pixel-fcn-route","operation":"move","beforeSHA256":"d764e90ca736ec69dae41ada1a8c4870b608d46d7dc46e61030c9bb13e328e9a","afterSHA256":"d764e90ca736ec69dae41ada1a8c4870b608d46d7dc46e61030c9bb13e328e9a","reason":"Same source body in authored semantic page"},{"oldId":"b-8a19ebde33325477520f0e94","targetId":"b-8a19ebde33325477520f0e94","page":"pixel-fcn-route","operation":"move","beforeSHA256":"40e51df4c794999a1c2b86aeb86dcbfba93bcbee0310637690744fac313d48c5","afterSHA256":"40e51df4c794999a1c2b86aeb86dcbfba93bcbee0310637690744fac313d48c5","reason":"Same source body in authored semantic page"},{"oldId":"b-a2bfa11ca57997b068f8e61a","targetId":"b-a2bfa11ca57997b068f8e61a","page":"pixel-fcn-route","operation":"move","beforeSHA256":"d0d7a14e2a82603ebe2f1ef958d50bde08aa07a6ffbf2a378807212760aa648f","afterSHA256":"d0d7a14e2a82603ebe2f1ef958d50bde08aa07a6ffbf2a378807212760aa648f","reason":"Same source body in authored semantic page"},{"oldId":"b-cb5be44fb4052ecd99d9472b","targetId":"b-cb5be44fb4052ecd99d9472b","page":"pixel-fcn-route","operation":"move","beforeSHA256":"51e5ab7e0ec16c70833117a3ce12a057283cc6bde9135affeb28a46d5509b290","afterSHA256":"51e5ab7e0ec16c70833117a3ce12a057283cc6bde9135affeb28a46d5509b290","reason":"Same source body in authored semantic page"},{"oldId":"b-e9f700afb2a077c0f6c2a024","targetId":"b-e9f700afb2a077c0f6c2a024","page":"pixel-fcn-route","operation":"move","beforeSHA256":"7bb7e83dd7b84bc279324f505b057ae90b458915ea105e690a8faaad0b78f279","afterSHA256":"7bb7e83dd7b84bc279324f505b057ae90b458915ea105e690a8faaad0b78f279","reason":"Same source body in authored semantic page"},{"oldId":"b-6214bb8a5291842657946be2","targetId":"b-6214bb8a5291842657946be2","page":"pixel-fcn-route","operation":"move","beforeSHA256":"6a9529ab1baed331c979e5be1de105172193e75ccf0b08ed4c28305472d225ea","afterSHA256":"6a9529ab1baed331c979e5be1de105172193e75ccf0b08ed4c28305472d225ea","reason":"Same source body in authored semantic page"},{"oldId":"b-3081385147a3b019bb5cb1f6","targetId":"b-3081385147a3b019bb5cb1f6","page":"pixel-fcn-route","operation":"move","beforeSHA256":"b2dbcbaa338e8a0c1cce0d2006f0c32340a94a455bb41b5c006c5b05b6a73e4f","afterSHA256":"b2dbcbaa338e8a0c1cce0d2006f0c32340a94a455bb41b5c006c5b05b6a73e4f","reason":"Same source body in authored semantic page"},{"oldId":"b-72ef42c09d3fd54e48f5a325","targetId":"b-72ef42c09d3fd54e48f5a325","page":"pixel-fcn-route","operation":"move","beforeSHA256":"fb49460b55fb3a2e3263ce964192e7e6eb01674373526ed187d209ae7e8580a8","afterSHA256":"fb49460b55fb3a2e3263ce964192e7e6eb01674373526ed187d209ae7e8580a8","reason":"Same source body in authored semantic page"},{"oldId":"b-0b271085254da2e4183908f1","targetId":"b-0b271085254da2e4183908f1","page":"pixel-fcn-route","operation":"move","beforeSHA256":"c1666f3770b7417119224b4288f60932500a799b89d6801efd77b82221aa0f7a","afterSHA256":"c1666f3770b7417119224b4288f60932500a799b89d6801efd77b82221aa0f7a","reason":"Same source body in authored semantic page"},{"oldId":"b-9a51884cbc7b0d70f6f1c515","targetId":"b-9a51884cbc7b0d70f6f1c515","page":"pixel-fcn-route","operation":"move","beforeSHA256":"99378d7450346cfed92c11f55449ec974c5d4a7c4d12e926404597c0125df371","afterSHA256":"99378d7450346cfed92c11f55449ec974c5d4a7c4d12e926404597c0125df371","reason":"Same source body in authored semantic page"},{"oldId":"b-1c45f1dfe33783f8422667dd","targetId":"b-1c45f1dfe33783f8422667dd","page":"pixel-supervision-contract","operation":"move","beforeSHA256":"f568ceb25e144397ab1a1ca1b9d741354418524c98eef10422cc026f6412eb5f","afterSHA256":"f568ceb25e144397ab1a1ca1b9d741354418524c98eef10422cc026f6412eb5f","reason":"Same source body in authored semantic page"},{"oldId":"b-ebfdaf3871efee314ae6bb41","targetId":"b-ebfdaf3871efee314ae6bb41","page":"pixel-supervision-contract","operation":"move","beforeSHA256":"eb0dfe6cebaefe68f519f69350ef606a2c687649403ed826851930d7434e55db","afterSHA256":"eb0dfe6cebaefe68f519f69350ef606a2c687649403ed826851930d7434e55db","reason":"Same source body in authored semantic page"},{"oldId":"b-eff7368c1e562de5e4c7bda9","targetId":"b-eff7368c1e562de5e4c7bda9","page":"pixel-supervision-contract","operation":"move","beforeSHA256":"ab380c1e0f9d1f36683516a23e6ae4f936078b69369c6e93f7772b9db9492422","afterSHA256":"ab380c1e0f9d1f36683516a23e6ae4f936078b69369c6e93f7772b9db9492422","reason":"Same source body in authored semantic page"},{"oldId":"b-81ef2b75713297fe6af8447a","targetId":"b-81ef2b75713297fe6af8447a","page":"pixel-supervision-contract","operation":"move","beforeSHA256":"a4c0dbd56af877d0b5e594f748f6f8d9603d840d6fc9a5838318be6e455ab3d5","afterSHA256":"a4c0dbd56af877d0b5e594f748f6f8d9603d840d6fc9a5838318be6e455ab3d5","reason":"Same source body in authored semantic page"},{"oldId":"b-5e979b40d5fa3e093a10c485","targetId":"b-5e979b40d5fa3e093a10c485","page":"pixel-supervision-contract","operation":"move","beforeSHA256":"fb67cbca43315249869971d864733783ccb53739e70b92e40c20d6581533bb82","afterSHA256":"fb67cbca43315249869971d864733783ccb53739e70b92e40c20d6581533bb82","reason":"Same source body in authored semantic page"},{"oldId":"b-1c03d2f3cf6c604af7872302","targetId":"b-1c03d2f3cf6c604af7872302","page":"pixel-supervision-contract","operation":"move","beforeSHA256":"75a26f3c334073d2322457d1f81001561d6c818c8f83b13b0f038ae1d77e5cc7","afterSHA256":"75a26f3c334073d2322457d1f81001561d6c818c8f83b13b0f038ae1d77e5cc7","reason":"Same source body in authored semantic page"},{"oldId":"b-728ba1710711069633cb474c","targetId":"b-728ba1710711069633cb474c","page":"pixel-lab11-baseline-evidence","operation":"move","beforeSHA256":"5bb25945392d1bafc8dd7d6177cf0bf83a1fca5b7d2f16e1def28c00bacba85f","afterSHA256":"5bb25945392d1bafc8dd7d6177cf0bf83a1fca5b7d2f16e1def28c00bacba85f","reason":"Same source body in authored semantic page"},{"oldId":"b-7cb21e34c6f920a5ef64b4f0","targetId":"b-7cb21e34c6f920a5ef64b4f0","page":"pixel-lab11-baseline-evidence","operation":"move","beforeSHA256":"6c0114296cc35b10699175b4c89039f6f9876e3185b83a7f02e7fa67ae9448e9","afterSHA256":"6c0114296cc35b10699175b4c89039f6f9876e3185b83a7f02e7fa67ae9448e9","reason":"Same source body in authored semantic page"},{"oldId":"b-ec9882427b0887ee99930f43","targetId":"b-ec9882427b0887ee99930f43","page":"pixel-lab11-baseline-evidence","operation":"move","beforeSHA256":"01b07d138cf5170bb6f10b5ab91b026d706baa6e021a922543b13d85cf3271ac","afterSHA256":"01b07d138cf5170bb6f10b5ab91b026d706baa6e021a922543b13d85cf3271ac","reason":"Same source body in authored semantic page"},{"oldId":"b-f8fd42219407d6858d415af0","targetId":"b-f8fd42219407d6858d415af0","page":"pixel-lab11-baseline-evidence","operation":"move","beforeSHA256":"53b208f148eb09e2dac3c533a2268e625a9dd513bee413b9a41b925e40631934","afterSHA256":"53b208f148eb09e2dac3c533a2268e625a9dd513bee413b9a41b925e40631934","reason":"Same source body in authored semantic page"},{"oldId":"b-c01b9fdea0abd68a152e49aa","targetId":"b-c01b9fdea0abd68a152e49aa","page":"pixel-lab11-baseline-evidence","operation":"move","beforeSHA256":"c31efb764a8e17b83caed97f41768f2bcebbb1a1964a23bcc8cbf6462d1b09a0","afterSHA256":"c31efb764a8e17b83caed97f41768f2bcebbb1a1964a23bcc8cbf6462d1b09a0","reason":"Same source body in authored semantic page"},{"oldId":"b-41c09ebcaca31c3ba48bc0b8","targetId":"b-41c09ebcaca31c3ba48bc0b8","page":"pixel-lab11-baseline-evidence","operation":"move","beforeSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","afterSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","reason":"Same source body in authored semantic page"},{"oldId":"b-eb316c507541bdcfeb765cd2","targetId":"b-eb316c507541bdcfeb765cd2","page":"pixel-lab11-baseline-evidence","operation":"move","beforeSHA256":"30703647a4a582d742b4724deb64e595de3208a7897216648bca33414cceabd0","afterSHA256":"30703647a4a582d742b4724deb64e595de3208a7897216648bca33414cceabd0","reason":"Same source body in authored semantic page"},{"oldId":"pixel-fcn-route-title","targetId":"b-dbf3ea24d4552e780309f390","operation":"retain-source-identity"},{"oldId":"pixel-supervision-contract-title","targetId":"b-72add4694bc3ba563062157e","operation":"retain-source-identity"},{"oldId":"pixel-supervision-contract-bridge","targetId":"b-5c3834c250104c6b15d1e336","operation":"retain-source-identity"},{"oldId":"pixel-lab11-baseline-evidence-title","targetId":"b-48874fbbaa74a13874598014","operation":"retain-source-identity"},{"oldId":"pixel-lab11-baseline-evidence-bridge","targetId":"b-8092bdd2f0e5e62669986b13","operation":"retain-source-identity"}],"blocks":[{"id":"b-dbf3ea24d4552e780309f390","type":"h2","scope":"pixel-fcn-route","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-dbf3ea24d4552e780309f390","content":"## 另起像素路线：整图直接预测类别地图","sha256":"2780a4f891f2d00017c51c872f1e474860e04e8fcef6cc11e75dc71a07c2ebdc","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"另起像素路线：整图直接预测类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-fccbf303d8f895e6e5afc7cf","prerequisites":[{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"}]},{"id":"b-fccbf303d8f895e6e5afc7cf","type":"h2","scope":"explanation","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-fccbf303d8f895e6e5afc7cf","content":"## 解释、算例与证据","sha256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"另起像素路线：整图直接预测类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-dbf3ea24d4552e780309f390","nextBlock":"b-53c3821247812cc98407a346","prerequisites":[{"id":"b-dbf3ea24d4552e780309f390","url":"#b-dbf3ea24d4552e780309f390","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"}]},{"id":"b-53c3821247812cc98407a346","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-53c3821247812cc98407a346","content":"# P36 · FCN：另一条路线，直接为像素作预测","sha256":"698f36893197b1d6e08b0bf5474b9348b819281814db1411c57d710d12f43df6","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"另起像素路线：整图直接预测类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-fccbf303d8f895e6e5afc7cf","nextBlock":"b-cd5deee9ce543a51a693e4b7","prerequisites":[{"id":"b-fccbf303d8f895e6e5afc7cf","url":"#b-fccbf303d8f895e6e5afc7cf","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"}]},{"id":"b-cd5deee9ce543a51a693e4b7","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-cd5deee9ce543a51a693e4b7","content":"FCN 从逐像素类别图出发，独立于区域检测路线。将 `[B,C,H,W]` logits 与 `[B,H,W]` 标签一一对应，理解分割损失和 ignore 像素。LAB11 从图像预测 mask；已知 mask 的缩放练习只检验采样，不检验视觉预测。","sha256":"183d41f9f2dd4315d6651bd668ce69b215479fec798b3e287bba7d446b7f4a31","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"另起像素路线：整图直接预测类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-53c3821247812cc98407a346","nextBlock":"b-b74111f2f662cd0dcfedc6a2","prerequisites":[{"id":"b-53c3821247812cc98407a346","url":"#b-53c3821247812cc98407a346","role":"section-title"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"}]},{"id":"b-b74111f2f662cd0dcfedc6a2","type":"figure","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-b74111f2f662cd0dcfedc6a2","content":"图内标签与数值：Lecture 03 课堂图解 P36；区域路线与像素路线，目标不同；R-CNN → Fast → Faster；一组对象：类别 + 框；FCN · 每位置类别分数；语义类别地图；Long · Shelhamer · Darrell，2014 预印本 / CVPR 2015；FCN 是密集像素预测路线，不是 Faster R-CNN 的下一代。\n\n区域检测路线与 FCN 像素路线分开；全图到空间特征再到语义 mask，署名 Long、Shelhamer、Darrell。\n\n- 区域路线与像素路线，目标不同\n- R-CNN → Fast → Faster\n- 一组对象：类别 + 框\n- FCN · 每位置类别分数\n- 语义类别地图\n- Long · Shelhamer · Darrell，2014 预印本 / CVPR 2015\n- FCN 是密集像素预测路线，不是 Faster R-CNN 的下一代。区域检测路线与 FCN 像素路线分开；全图到空间特征再到语义 mask，署名 Long、Shelhamer、Darrell。 [查看对应图解状态](/#/l3-r2-p36/0)[查看完整图解](/course/figures/l3-r2-p36-0.html)","sha256":"fafc5dc0746074c8919d638becd9179587704ce91ea6431b2795bd9bae0eb43f","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"另起像素路线：整图直接预测类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"区域检测路线与 FCN 像素路线分开；全图到空间特征再到语义 mask，署名 Long、Shelhamer、Darrell。 查看对应图解状态","labels":["区域路线与像素路线，目标不同","R-CNN → Fast → Faster","一组对象：类别 + 框","FCN · 每位置类别分数","语义类别地图","Long · Shelhamer · Darrell，2014 预印本 / CVPR 2015","FCN 是密集像素预测路线，不是 Faster R-CNN 的下一代。"],"description":"区域检测路线与 FCN 像素路线分开；全图到空间特征再到语义 mask，署名 Long、Shelhamer、Darrell。区域路线与像素路线，目标不同R-CNN → Fast → Faster一组对象：类别 + 框FCN · 每位置类别分数语义类别地图Long · Shelhamer · Darrell，2014 预印本 / CVPR 2015FCN 是密集像素预测路线，不是 Faster R-CNN 的下一代。","evidenceIdentity":"区域检测路线与 FCN 像素路线分开；全图到空间特征再到语义 mask，署名 Long、Shelhamer、Darrell。 查看对应图解状态","images":[{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""},{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""}]},"previousBlock":"b-cd5deee9ce543a51a693e4b7","nextBlock":"b-662c1d2a0499d3ca223cf2cf","prerequisites":[{"id":"b-53c3821247812cc98407a346","url":"#b-53c3821247812cc98407a346","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-662c1d2a0499d3ca223cf2cf","type":"h2","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-662c1d2a0499d3ca223cf2cf","content":"## 课堂讲述","sha256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"另起像素路线：整图直接预测类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b74111f2f662cd0dcfedc6a2","nextBlock":"b-1c949701f78f7f0e33c9873e","prerequisites":[{"id":"b-53c3821247812cc98407a346","url":"#b-53c3821247812cc98407a346","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"}]},{"id":"b-1c949701f78f7f0e33c9873e","type":"p","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-1c949701f78f7f0e33c9873e","content":"前面以区域为单位组织检测。另一个问题是：能不能直接为整张图上的每个位置给出类别？这就是语义分割的像素预测路线。\nFCN 把分类网络中的计算改造成适合密集输出的形式，让空间位置对应到类别预测。它不是 Fast 或 Faster R-CNN 后面的自然“下一代”，而是有独立目标的一条路线。Long、Shelhamer 和 Darrell 的工作，要在这个问题里理解。\n请把两种输出摆在一起看：检测输出一组对象，语义分割输出一张类别地图。如果我们的科研任务关心组织、地表或材料区域的归属，后者就是直接的任务接口。接下来真正困难的是：网络深层的空间网格通常已经变粗，怎样交付像素级结果？","sha256":"cd13f386087a944b59dbd751e7acf3e67a0cea054518c1212856523a6b549dc2","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"另起像素路线：整图直接预测类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-662c1d2a0499d3ca223cf2cf","nextBlock":"b-283c4597a38d3776bae3455f","prerequisites":[{"id":"b-662c1d2a0499d3ca223cf2cf","url":"#b-662c1d2a0499d3ca223cf2cf","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"}]},{"id":"b-283c4597a38d3776bae3455f","type":"h3","scope":"fcn-independent-route","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-283c4597a38d3776bae3455f","content":"### FCN 直接学习一张类别地图","sha256":"689bfff2fcfc0e090d7c51735f916750b81432aa009d448f5600b865c9742259","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"FCN 直接学习一张类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":"fcn-independent-route","formulas":[],"previousBlock":"b-1c949701f78f7f0e33c9873e","nextBlock":"b-5354e2cc50e944a5cfeaef6f","prerequisites":[{"id":"b-662c1d2a0499d3ca223cf2cf","url":"#b-662c1d2a0499d3ca223cf2cf","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"},{"id":"b-5354e2cc50e944a5cfeaef6f","url":"#b-5354e2cc50e944a5cfeaef6f","relation":"same-reasoning"},{"id":"b-8a19ebde33325477520f0e94","url":"#b-8a19ebde33325477520f0e94","relation":"same-reasoning"},{"id":"b-a2bfa11ca57997b068f8e61a","url":"#b-a2bfa11ca57997b068f8e61a","relation":"same-reasoning"},{"id":"b-cb5be44fb4052ecd99d9472b","url":"#b-cb5be44fb4052ecd99d9472b","relation":"same-reasoning"}]},{"id":"b-5354e2cc50e944a5cfeaef6f","type":"p","scope":"fcn-independent-route","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-5354e2cc50e944a5cfeaef6f","content":"如果用途是给组织切片的每个像素指定类别，先列一组外接框可能并不是自然起点。FCN 从空间网格出发：把分类网络改成可对不同输入大小输出空间分数图的计算，接受逐像素标签训练，再恢复到目标分辨率。","sha256":"d764e90ca736ec69dae41ada1a8c4870b608d46d7dc46e61030c9bb13e328e9a","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"FCN 直接学习一张类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":"fcn-independent-route","formulas":[],"previousBlock":"b-283c4597a38d3776bae3455f","nextBlock":"b-8a19ebde33325477520f0e94","prerequisites":[{"id":"b-283c4597a38d3776bae3455f","url":"#b-283c4597a38d3776bae3455f","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"},{"id":"b-283c4597a38d3776bae3455f","url":"#b-283c4597a38d3776bae3455f","relation":"same-reasoning"},{"id":"b-8a19ebde33325477520f0e94","url":"#b-8a19ebde33325477520f0e94","relation":"same-reasoning"},{"id":"b-a2bfa11ca57997b068f8e61a","url":"#b-a2bfa11ca57997b068f8e61a","relation":"same-reasoning"},{"id":"b-cb5be44fb4052ecd99d9472b","url":"#b-cb5be44fb4052ecd99d9472b","relation":"same-reasoning"}]},{"id":"b-8a19ebde33325477520f0e94","type":"p","scope":"fcn-independent-route","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-8a19ebde33325477520f0e94","content":"一个 C 类分割模型可以交付 `[B,C,H,W]` logits，每个像素有 C 个实数；目标 `[B,H,W]` 保存该像素的整数类别。类别维 softmax 决定“同一位置是哪类”，空间维则保留“这个判断在哪里”。如果把所有空间位置先平均成一个类别向量，就改了任务。","sha256":"40e51df4c794999a1c2b86aeb86dcbfba93bcbee0310637690744fac313d48c5","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"FCN 直接学习一张类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":"fcn-independent-route","formulas":[],"previousBlock":"b-5354e2cc50e944a5cfeaef6f","nextBlock":"b-a2bfa11ca57997b068f8e61a","prerequisites":[{"id":"b-283c4597a38d3776bae3455f","url":"#b-283c4597a38d3776bae3455f","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"},{"id":"b-283c4597a38d3776bae3455f","url":"#b-283c4597a38d3776bae3455f","relation":"same-reasoning"},{"id":"b-5354e2cc50e944a5cfeaef6f","url":"#b-5354e2cc50e944a5cfeaef6f","relation":"same-reasoning"},{"id":"b-a2bfa11ca57997b068f8e61a","url":"#b-a2bfa11ca57997b068f8e61a","relation":"same-reasoning"},{"id":"b-cb5be44fb4052ecd99d9472b","url":"#b-cb5be44fb4052ecd99d9472b","relation":"same-reasoning"}]},{"id":"b-a2bfa11ca57997b068f8e61a","type":"p","scope":"fcn-independent-route","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-a2bfa11ca57997b068f8e61a","content":"FCN 与区域检测是两条并行问题线：前者组织像素类别，后者组织对象集合。两条线都能利用分类预训练，却不能把 FCN 描述为 Faster R-CNN 的后继版本。原始 2014/2015 FCN 摘要报告 VOC 2012 的 mean IU 为 62.2%；2016 扩展版本有不同数字，引用时需区分版本。","sha256":"d0d7a14e2a82603ebe2f1ef958d50bde08aa07a6ffbf2a378807212760aa648f","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"FCN 直接学习一张类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":"fcn-independent-route","formulas":[],"previousBlock":"b-8a19ebde33325477520f0e94","nextBlock":"b-cb5be44fb4052ecd99d9472b","prerequisites":[{"id":"b-283c4597a38d3776bae3455f","url":"#b-283c4597a38d3776bae3455f","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"},{"id":"b-283c4597a38d3776bae3455f","url":"#b-283c4597a38d3776bae3455f","relation":"same-reasoning"},{"id":"b-5354e2cc50e944a5cfeaef6f","url":"#b-5354e2cc50e944a5cfeaef6f","relation":"same-reasoning"},{"id":"b-8a19ebde33325477520f0e94","url":"#b-8a19ebde33325477520f0e94","relation":"same-reasoning"},{"id":"b-cb5be44fb4052ecd99d9472b","url":"#b-cb5be44fb4052ecd99d9472b","relation":"same-reasoning"}]},{"id":"b-cb5be44fb4052ecd99d9472b","type":"p","scope":"fcn-independent-route","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-cb5be44fb4052ecd99d9472b","content":"最值得带回科研任务的是这个接口改变：表示能迁移，并不意味着原输出头能直接交付新答案。标注的单位变成像素，损失与评价也须随之改变。","sha256":"51e5ab7e0ec16c70833117a3ce12a057283cc6bde9135affeb28a46d5509b290","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"FCN 直接学习一张类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":"fcn-independent-route","formulas":[],"previousBlock":"b-a2bfa11ca57997b068f8e61a","nextBlock":"b-186a339cfb17544ffd045d7f","prerequisites":[{"id":"b-283c4597a38d3776bae3455f","url":"#b-283c4597a38d3776bae3455f","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"},{"id":"b-283c4597a38d3776bae3455f","url":"#b-283c4597a38d3776bae3455f","relation":"same-reasoning"},{"id":"b-5354e2cc50e944a5cfeaef6f","url":"#b-5354e2cc50e944a5cfeaef6f","relation":"same-reasoning"},{"id":"b-8a19ebde33325477520f0e94","url":"#b-8a19ebde33325477520f0e94","relation":"same-reasoning"},{"id":"b-a2bfa11ca57997b068f8e61a","url":"#b-a2bfa11ca57997b068f8e61a","relation":"same-reasoning"}]},{"id":"b-186a339cfb17544ffd045d7f","type":"h4","scope":"pixel-fcn-route","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-186a339cfb17544ffd045d7f","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":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"另起像素路线：整图直接预测类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-cb5be44fb4052ecd99d9472b","nextBlock":"b-e9f700afb2a077c0f6c2a024","prerequisites":[{"id":"b-283c4597a38d3776bae3455f","url":"#b-283c4597a38d3776bae3455f","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"}]},{"id":"b-e9f700afb2a077c0f6c2a024","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-e9f700afb2a077c0f6c2a024","content":"FCN 是 Long、Shelhamer、Darrell 的独立密集预测路线。不要把课堂先讲 Faster 的顺序理解成 FCN 历史上后继于 Faster。","sha256":"7bb7e83dd7b84bc279324f505b057ae90b458915ea105e690a8faaad0b78f279","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"另起像素路线：整图直接预测类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-186a339cfb17544ffd045d7f","nextBlock":"b-fd2285691c0f6492cf1beb21","prerequisites":[{"id":"b-186a339cfb17544ffd045d7f","url":"#b-186a339cfb17544ffd045d7f","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"}]},{"id":"b-fd2285691c0f6492cf1beb21","type":"h4","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-fd2285691c0f6492cf1beb21","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":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"另起像素路线：整图直接预测类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e9f700afb2a077c0f6c2a024","nextBlock":"b-6214bb8a5291842657946be2","prerequisites":[{"id":"b-186a339cfb17544ffd045d7f","url":"#b-186a339cfb17544ffd045d7f","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"}]},{"id":"b-6214bb8a5291842657946be2","type":"ul","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-6214bb8a5291842657946be2","content":"- [每个位置，都要一个类别](/supplement.html?return=l3-r2-p36/0#/l3-fcn/0) 本页说明\n- [FCN 是独立的密集预测路线｜问题从哪里出现](/supplement.html?return=l3-r2-p36/0#/l3-fcn-dense-problem/0) 本页说明\n- [FCN 是独立的密集预测路线｜核心改动怎样工作](/supplement.html?return=l3-r2-p36/0#/l3-fcn-dense-mechanism/0) 本页说明\n- [FCN 是独立的密集预测路线｜什么结论还不能推出](/supplement.html?return=l3-r2-p36/0#/l3-fcn-dense-boundary/0) 本页说明","sha256":"6a9529ab1baed331c979e5be1de105172193e75ccf0b08ed4c28305472d225ea","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"另起像素路线：整图直接预测类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-fd2285691c0f6492cf1beb21","nextBlock":"b-3081385147a3b019bb5cb1f6","prerequisites":[{"id":"b-fd2285691c0f6492cf1beb21","url":"#b-fd2285691c0f6492cf1beb21","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"}]},{"id":"b-3081385147a3b019bb5cb1f6","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-3081385147a3b019bb5cb1f6","content":"- [每个位置，都要一个类别](/supplement.html?return=l3-r2-p36/0#/l3-fcn/0) 本页说明","sha256":"b2dbcbaa338e8a0c1cce0d2006f0c32340a94a455bb41b5c006c5b05b6a73e4f","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"另起像素路线：整图直接预测类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6214bb8a5291842657946be2","nextBlock":"b-72ef42c09d3fd54e48f5a325","prerequisites":[{"id":"b-fd2285691c0f6492cf1beb21","url":"#b-fd2285691c0f6492cf1beb21","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"}]},{"id":"b-72ef42c09d3fd54e48f5a325","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-72ef42c09d3fd54e48f5a325","content":"- [FCN 是独立的密集预测路线｜问题从哪里出现](/supplement.html?return=l3-r2-p36/0#/l3-fcn-dense-problem/0) 本页说明","sha256":"fb49460b55fb3a2e3263ce964192e7e6eb01674373526ed187d209ae7e8580a8","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"另起像素路线：整图直接预测类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3081385147a3b019bb5cb1f6","nextBlock":"b-0b271085254da2e4183908f1","prerequisites":[{"id":"b-fd2285691c0f6492cf1beb21","url":"#b-fd2285691c0f6492cf1beb21","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"}]},{"id":"b-0b271085254da2e4183908f1","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-0b271085254da2e4183908f1","content":"- [FCN 是独立的密集预测路线｜核心改动怎样工作](/supplement.html?return=l3-r2-p36/0#/l3-fcn-dense-mechanism/0) 本页说明","sha256":"c1666f3770b7417119224b4288f60932500a799b89d6801efd77b82221aa0f7a","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"另起像素路线：整图直接预测类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-72ef42c09d3fd54e48f5a325","nextBlock":"b-9a51884cbc7b0d70f6f1c515","prerequisites":[{"id":"b-fd2285691c0f6492cf1beb21","url":"#b-fd2285691c0f6492cf1beb21","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"}]},{"id":"b-9a51884cbc7b0d70f6f1c515","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-9a51884cbc7b0d70f6f1c515","content":"- [FCN 是独立的密集预测路线｜什么结论还不能推出](/supplement.html?return=l3-r2-p36/0#/l3-fcn-dense-boundary/0) 本页说明","sha256":"99378d7450346cfed92c11f55449ec974c5d4a7c4d12e926404597c0125df371","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-fcn-route","title":"另起像素路线：整图直接预测类别地图","number":1},"sectionTitle":"另起像素路线：整图直接预测类别地图","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-0b271085254da2e4183908f1","nextBlock":null,"prerequisites":[{"id":"b-fd2285691c0f6492cf1beb21","url":"#b-fd2285691c0f6492cf1beb21","role":"section-title"},{"id":"b-cd5deee9ce543a51a693e4b7","url":"#b-cd5deee9ce543a51a693e4b7","role":"problem-context"}],"relatedBlocks":[{"id":"b-b74111f2f662cd0dcfedc6a2","url":"#b-b74111f2f662cd0dcfedc6a2","relation":"figure-and-explanation"}]},{"id":"b-72add4694bc3ba563062157e","type":"h2","scope":"pixel-supervision-contract","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-72add4694bc3ba563062157e","content":"## 让 logits、标签和有效像素一一对应","sha256":"3eda454d68dfcad68c252d87cbdccb22443c6dd2d4bcb24fc154d1b31c67a8a4","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-supervision-contract","title":"让 logits、标签和有效像素一一对应","number":2},"sectionTitle":"让 logits、标签和有效像素一一对应","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36","l3-r2-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-5c3834c250104c6b15d1e336","prerequisites":[{"id":"b-5c3834c250104c6b15d1e336","url":"#b-5c3834c250104c6b15d1e336","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-5c3834c250104c6b15d1e336","type":"p","scope":"pixel-supervision-contract","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-5c3834c250104c6b15d1e336","content":"先建立像素坐标与类别标签的对应，再解释损失和忽略位置。","sha256":"e3ab353f84cf926eefc2ac9bf7c7e6d044a5706d26ac2fabe7332bd6936e0ff5","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-supervision-contract","title":"让 logits、标签和有效像素一一对应","number":2},"sectionTitle":"让 logits、标签和有效像素一一对应","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36","l3-r2-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-72add4694bc3ba563062157e","nextBlock":"b-1c45f1dfe33783f8422667dd","prerequisites":[{"id":"b-72add4694bc3ba563062157e","url":"#b-72add4694bc3ba563062157e","role":"section-title"}],"relatedBlocks":[]},{"id":"b-1c45f1dfe33783f8422667dd","type":"h3","scope":"pixel-cross-entropy","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-1c45f1dfe33783f8422667dd","content":"### 每个有效像素提供一项监督","sha256":"f568ceb25e144397ab1a1ca1b9d741354418524c98eef10422cc026f6412eb5f","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-supervision-contract","title":"让 logits、标签和有效像素一一对应","number":2},"sectionTitle":"每个有效像素提供一项监督","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36","l3-r2-p37"],"knowledgeBlock":"pixel-cross-entropy","formulas":[],"previousBlock":"b-5c3834c250104c6b15d1e336","nextBlock":"b-ebfdaf3871efee314ae6bb41","prerequisites":[{"id":"b-72add4694bc3ba563062157e","url":"#b-72add4694bc3ba563062157e","role":"section-title"},{"id":"b-5c3834c250104c6b15d1e336","url":"#b-5c3834c250104c6b15d1e336","role":"problem-context"}],"relatedBlocks":[{"id":"b-ebfdaf3871efee314ae6bb41","url":"#b-ebfdaf3871efee314ae6bb41","relation":"same-reasoning"},{"id":"b-eff7368c1e562de5e4c7bda9","url":"#b-eff7368c1e562de5e4c7bda9","relation":"same-reasoning"},{"id":"b-81ef2b75713297fe6af8447a","url":"#b-81ef2b75713297fe6af8447a","relation":"same-reasoning"},{"id":"b-5e979b40d5fa3e093a10c485","url":"#b-5e979b40d5fa3e093a10c485","relation":"same-reasoning"},{"id":"b-1c03d2f3cf6c604af7872302","url":"#b-1c03d2f3cf6c604af7872302","relation":"same-reasoning"}]},{"id":"b-ebfdaf3871efee314ae6bb41","type":"p","scope":"pixel-cross-entropy","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-ebfdaf3871efee314ae6bb41","content":"对目标类别为 $y_{uv}$ 的像素，交叉熵是 $-\\log p_{y_{uv},u,v}$。若有效像素集合为 Ω，平均损失写成","sha256":"eb0dfe6cebaefe68f519f69350ef606a2c687649403ed826851930d7434e55db","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-supervision-contract","title":"让 logits、标签和有效像素一一对应","number":2},"sectionTitle":"每个有效像素提供一项监督","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36","l3-r2-p37"],"knowledgeBlock":"pixel-cross-entropy","formulas":[{"latex":"y_{uv}","definitionContext":"对目标类别为 $y_{uv}$ 的像素，交叉熵是 $-\\log p_{y_{uv},u,v}$。若有效像素集合为 Ω，平均损失写成\n$\nL=-\\frac{1}{|\\Omega|}\\sum_{(u,v)\\in\\Omega}\\log p_{y_{uv},u,v}.\n$\n三类 logits `[B,3,H,W]` 与整数目标 `[B,H,W]` 相配；不要把类别 ID 当成连续数直接算回归平方误差。二类分割也可只预测一个前景 logit，用 sigmoid 与二元交叉熵，这时标签与输出接口跟三类方案不同。\n未标注位置常用 ignore=255。它表示“不在这里计算监督和指标”，不表示第四类，也不等于背景。LAB11 的损失和混淆矩阵都排除这些位置；否则模型可能因为迎合不存在的背景标签而产生错误梯度。\n图像翻转、裁剪或缩放时，目标地图必须接受同一几何变换。RGB 图能用连续插值，离散 mask 通常用最近邻；把类 1 与类 3 插出“类 2”没有语义依据。先保证一处图像像素仍对应同一处目标，训练曲线才有解释价值。"},{"latex":"-\\log p_{y_{uv},u,v}","definitionContext":"对目标类别为 $y_{uv}$ 的像素，交叉熵是 $-\\log p_{y_{uv},u,v}$。若有效像素集合为 Ω，平均损失写成\n$\nL=-\\frac{1}{|\\Omega|}\\sum_{(u,v)\\in\\Omega}\\log p_{y_{uv},u,v}.\n$\n三类 logits `[B,3,H,W]` 与整数目标 `[B,H,W]` 相配；不要把类别 ID 当成连续数直接算回归平方误差。二类分割也可只预测一个前景 logit，用 sigmoid 与二元交叉熵，这时标签与输出接口跟三类方案不同。\n未标注位置常用 ignore=255。它表示“不在这里计算监督和指标”，不表示第四类，也不等于背景。LAB11 的损失和混淆矩阵都排除这些位置；否则模型可能因为迎合不存在的背景标签而产生错误梯度。\n图像翻转、裁剪或缩放时，目标地图必须接受同一几何变换。RGB 图能用连续插值，离散 mask 通常用最近邻；把类 1 与类 3 插出“类 2”没有语义依据。先保证一处图像像素仍对应同一处目标，训练曲线才有解释价值。"}],"previousBlock":"b-1c45f1dfe33783f8422667dd","nextBlock":"b-eff7368c1e562de5e4c7bda9","prerequisites":[{"id":"b-1c45f1dfe33783f8422667dd","url":"#b-1c45f1dfe33783f8422667dd","role":"section-title"},{"id":"b-5c3834c250104c6b15d1e336","url":"#b-5c3834c250104c6b15d1e336","role":"problem-context"}],"relatedBlocks":[{"id":"b-1c45f1dfe33783f8422667dd","url":"#b-1c45f1dfe33783f8422667dd","relation":"same-reasoning"},{"id":"b-eff7368c1e562de5e4c7bda9","url":"#b-eff7368c1e562de5e4c7bda9","relation":"same-reasoning"},{"id":"b-81ef2b75713297fe6af8447a","url":"#b-81ef2b75713297fe6af8447a","relation":"same-reasoning"},{"id":"b-5e979b40d5fa3e093a10c485","url":"#b-5e979b40d5fa3e093a10c485","relation":"same-reasoning"},{"id":"b-1c03d2f3cf6c604af7872302","url":"#b-1c03d2f3cf6c604af7872302","relation":"same-reasoning"}]},{"id":"b-eff7368c1e562de5e4c7bda9","type":"p","scope":"pixel-cross-entropy","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-eff7368c1e562de5e4c7bda9","content":"$\nL=-\\frac{1}{|\\Omega|}\\sum_{(u,v)\\in\\Omega}\\log p_{y_{uv},u,v}.\n$","sha256":"ab380c1e0f9d1f36683516a23e6ae4f936078b69369c6e93f7772b9db9492422","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-supervision-contract","title":"让 logits、标签和有效像素一一对应","number":2},"sectionTitle":"每个有效像素提供一项监督","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36","l3-r2-p37"],"knowledgeBlock":"pixel-cross-entropy","formulas":[{"latex":"\nL=-\\frac{1}{|\\Omega|}\\sum_{(u,v)\\in\\Omega}\\log p_{y_{uv},u,v}.\n","definitionContext":"对目标类别为 $y_{uv}$ 的像素，交叉熵是 $-\\log p_{y_{uv},u,v}$。若有效像素集合为 Ω，平均损失写成\n$\nL=-\\frac{1}{|\\Omega|}\\sum_{(u,v)\\in\\Omega}\\log p_{y_{uv},u,v}.\n$\n三类 logits `[B,3,H,W]` 与整数目标 `[B,H,W]` 相配；不要把类别 ID 当成连续数直接算回归平方误差。二类分割也可只预测一个前景 logit，用 sigmoid 与二元交叉熵，这时标签与输出接口跟三类方案不同。\n未标注位置常用 ignore=255。它表示“不在这里计算监督和指标”，不表示第四类，也不等于背景。LAB11 的损失和混淆矩阵都排除这些位置；否则模型可能因为迎合不存在的背景标签而产生错误梯度。\n图像翻转、裁剪或缩放时，目标地图必须接受同一几何变换。RGB 图能用连续插值，离散 mask 通常用最近邻；把类 1 与类 3 插出“类 2”没有语义依据。先保证一处图像像素仍对应同一处目标，训练曲线才有解释价值。"}],"previousBlock":"b-ebfdaf3871efee314ae6bb41","nextBlock":"b-81ef2b75713297fe6af8447a","prerequisites":[{"id":"b-1c45f1dfe33783f8422667dd","url":"#b-1c45f1dfe33783f8422667dd","role":"section-title"},{"id":"b-5c3834c250104c6b15d1e336","url":"#b-5c3834c250104c6b15d1e336","role":"problem-context"}],"relatedBlocks":[{"id":"b-1c45f1dfe33783f8422667dd","url":"#b-1c45f1dfe33783f8422667dd","relation":"same-reasoning"},{"id":"b-ebfdaf3871efee314ae6bb41","url":"#b-ebfdaf3871efee314ae6bb41","relation":"same-reasoning"},{"id":"b-81ef2b75713297fe6af8447a","url":"#b-81ef2b75713297fe6af8447a","relation":"same-reasoning"},{"id":"b-5e979b40d5fa3e093a10c485","url":"#b-5e979b40d5fa3e093a10c485","relation":"same-reasoning"},{"id":"b-1c03d2f3cf6c604af7872302","url":"#b-1c03d2f3cf6c604af7872302","relation":"same-reasoning"}]},{"id":"b-81ef2b75713297fe6af8447a","type":"p","scope":"pixel-cross-entropy","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-81ef2b75713297fe6af8447a","content":"三类 logits `[B,3,H,W]` 与整数目标 `[B,H,W]` 相配；不要把类别 ID 当成连续数直接算回归平方误差。二类分割也可只预测一个前景 logit，用 sigmoid 与二元交叉熵，这时标签与输出接口跟三类方案不同。","sha256":"a4c0dbd56af877d0b5e594f748f6f8d9603d840d6fc9a5838318be6e455ab3d5","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-supervision-contract","title":"让 logits、标签和有效像素一一对应","number":2},"sectionTitle":"每个有效像素提供一项监督","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36","l3-r2-p37"],"knowledgeBlock":"pixel-cross-entropy","formulas":[],"previousBlock":"b-eff7368c1e562de5e4c7bda9","nextBlock":"b-5e979b40d5fa3e093a10c485","prerequisites":[{"id":"b-1c45f1dfe33783f8422667dd","url":"#b-1c45f1dfe33783f8422667dd","role":"section-title"},{"id":"b-5c3834c250104c6b15d1e336","url":"#b-5c3834c250104c6b15d1e336","role":"problem-context"}],"relatedBlocks":[{"id":"b-1c45f1dfe33783f8422667dd","url":"#b-1c45f1dfe33783f8422667dd","relation":"same-reasoning"},{"id":"b-ebfdaf3871efee314ae6bb41","url":"#b-ebfdaf3871efee314ae6bb41","relation":"same-reasoning"},{"id":"b-eff7368c1e562de5e4c7bda9","url":"#b-eff7368c1e562de5e4c7bda9","relation":"same-reasoning"},{"id":"b-5e979b40d5fa3e093a10c485","url":"#b-5e979b40d5fa3e093a10c485","relation":"same-reasoning"},{"id":"b-1c03d2f3cf6c604af7872302","url":"#b-1c03d2f3cf6c604af7872302","relation":"same-reasoning"}]},{"id":"b-5e979b40d5fa3e093a10c485","type":"p","scope":"pixel-cross-entropy","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-5e979b40d5fa3e093a10c485","content":"未标注位置常用 ignore=255。它表示“不在这里计算监督和指标”，不表示第四类，也不等于背景。LAB11 的损失和混淆矩阵都排除这些位置；否则模型可能因为迎合不存在的背景标签而产生错误梯度。","sha256":"fb67cbca43315249869971d864733783ccb53739e70b92e40c20d6581533bb82","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-supervision-contract","title":"让 logits、标签和有效像素一一对应","number":2},"sectionTitle":"每个有效像素提供一项监督","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36","l3-r2-p37"],"knowledgeBlock":"pixel-cross-entropy","formulas":[],"previousBlock":"b-81ef2b75713297fe6af8447a","nextBlock":"b-1c03d2f3cf6c604af7872302","prerequisites":[{"id":"b-1c45f1dfe33783f8422667dd","url":"#b-1c45f1dfe33783f8422667dd","role":"section-title"},{"id":"b-5c3834c250104c6b15d1e336","url":"#b-5c3834c250104c6b15d1e336","role":"problem-context"}],"relatedBlocks":[{"id":"b-1c45f1dfe33783f8422667dd","url":"#b-1c45f1dfe33783f8422667dd","relation":"same-reasoning"},{"id":"b-ebfdaf3871efee314ae6bb41","url":"#b-ebfdaf3871efee314ae6bb41","relation":"same-reasoning"},{"id":"b-eff7368c1e562de5e4c7bda9","url":"#b-eff7368c1e562de5e4c7bda9","relation":"same-reasoning"},{"id":"b-81ef2b75713297fe6af8447a","url":"#b-81ef2b75713297fe6af8447a","relation":"same-reasoning"},{"id":"b-1c03d2f3cf6c604af7872302","url":"#b-1c03d2f3cf6c604af7872302","relation":"same-reasoning"}]},{"id":"b-1c03d2f3cf6c604af7872302","type":"p","scope":"pixel-cross-entropy","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-1c03d2f3cf6c604af7872302","content":"图像翻转、裁剪或缩放时，目标地图必须接受同一几何变换。RGB 图能用连续插值，离散 mask 通常用最近邻；把类 1 与类 3 插出“类 2”没有语义依据。先保证一处图像像素仍对应同一处目标，训练曲线才有解释价值。","sha256":"75a26f3c334073d2322457d1f81001561d6c818c8f83b13b0f038ae1d77e5cc7","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-prediction-objective","title":"怎样直接学习一张像素类别地图？"},"page":{"id":"pixel-supervision-contract","title":"让 logits、标签和有效像素一一对应","number":2},"sectionTitle":"每个有效像素提供一项监督","currentQuestion":"从区域对象集合转向逐像素目标，监督和指标如何对应？","sourceSlides":["l3-r2-p36","l3-r2-p37"],"knowledgeBlock":"pixel-cross-entropy","formulas":[],"previousBlock":"b-5e979b40d5fa3e093a10c485","nextBlock":null,"prerequisites":[{"id":"b-1c45f1dfe33783f8422667dd","url":"#b-1c45f1dfe33783f8422667dd","role":"section-title"},{"id":"b-5c3834c250104c6b15d1e336","url":"#b-5c3834c250104c6b15d1e336","role":"problem-context"}],"relatedBlocks":[{"id":"b-1c45f1dfe33783f8422667dd","url":"#b-1c45f1dfe33783f8422667dd","relation":"same-reasoning"},{"id":"b-ebfdaf3871efee314ae6bb41","url":"#b-ebfdaf3871efee314ae6bb41","relation":"same-reasoning"},{"id":"b-eff7368c1e562de5e4c7bda9","url":"#b-eff7368c1e562de5e4c7bda9","relation":"same-reasoning"},{"id":"b-81ef2b75713297fe6af8447a","url":"#b-81ef2b75713297fe6af8447a","relation":"same-reasoning"},{"id":"b-5e979b40d5fa3e093a10c485","url":"#b-5e979b40d5fa3e093a10c485","relation":"same-reasoning"}]},{"id":"b-48874fbbaa74a13874598014","type":"h2","scope":"pixel-lab11-baseline-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-48874fbbaa74a13874598014","content":"## 进入 LAB11：保留简单基线再看条件变化","sha256":"9fdbbbab239583f400bebb8199d90378f507b07ffa984e5912f9cf177057d082","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"pixel-lab11-baseline-evidence","title":"进入 LAB11：保留简单基线再看条件变化","number":3},"sectionTitle":"进入 LAB11：保留简单基线再看条件变化","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p36","l3-r2-p38"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-8092bdd2f0e5e62669986b13","prerequisites":[{"id":"b-8092bdd2f0e5e62669986b13","url":"#b-8092bdd2f0e5e62669986b13","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-8092bdd2f0e5e62669986b13","type":"p","scope":"pixel-lab11-baseline-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-8092bdd2f0e5e62669986b13","content":"有了机制假设，再用已有同条件结果判断额外结构的实际作用。","sha256":"4d491e50f316b4723176826810616bb5e72e3bcaed1050a912755dd8692c1485","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"pixel-lab11-baseline-evidence","title":"进入 LAB11：保留简单基线再看条件变化","number":3},"sectionTitle":"进入 LAB11：保留简单基线再看条件变化","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p36","l3-r2-p38"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-48874fbbaa74a13874598014","nextBlock":"b-728ba1710711069633cb474c","prerequisites":[{"id":"b-48874fbbaa74a13874598014","url":"#b-48874fbbaa74a13874598014","role":"section-title"}],"relatedBlocks":[]},{"id":"b-728ba1710711069633cb474c","type":"h3","scope":"pixel-model-positive-result","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-728ba1710711069633cb474c","content":"### 同一组分割实验里，简单模型为何值得留下","sha256":"5bb25945392d1bafc8dd7d6177cf0bf83a1fca5b7d2f16e1def28c00bacba85f","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"pixel-lab11-baseline-evidence","title":"进入 LAB11：保留简单基线再看条件变化","number":3},"sectionTitle":"同一组分割实验里，简单模型为何值得留下","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p36","l3-r2-p38"],"knowledgeBlock":"pixel-model-positive-result","formulas":[],"previousBlock":"b-8092bdd2f0e5e62669986b13","nextBlock":"b-7cb21e34c6f920a5ef64b4f0","prerequisites":[{"id":"b-48874fbbaa74a13874598014","url":"#b-48874fbbaa74a13874598014","role":"section-title"},{"id":"b-8092bdd2f0e5e62669986b13","url":"#b-8092bdd2f0e5e62669986b13","role":"problem-context"}],"relatedBlocks":[{"id":"b-7cb21e34c6f920a5ef64b4f0","url":"#b-7cb21e34c6f920a5ef64b4f0","relation":"same-reasoning"},{"id":"b-ec9882427b0887ee99930f43","url":"#b-ec9882427b0887ee99930f43","relation":"same-reasoning"},{"id":"b-f8fd42219407d6858d415af0","url":"#b-f8fd42219407d6858d415af0","relation":"same-reasoning"},{"id":"b-c01b9fdea0abd68a152e49aa","url":"#b-c01b9fdea0abd68a152e49aa","relation":"same-reasoning"}]},{"id":"b-7cb21e34c6f920a5ef64b4f0","type":"p","scope":"pixel-model-positive-result","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-7cb21e34c6f920a5ef64b4f0","content":"当前 LAB11 同时训练逐像素 RGB 线性读出和小型 FCN。输入 48×48 RGB，标签是背景、红矩形、蓝椭圆；384 张训练图与 128 张保留图按场景种子分开。FCN 用粗分数与细分支融合，RGB 基线则只用 `Conv2d(3,3,1)`，共 12 个参数，不能看邻域。","sha256":"6c0114296cc35b10699175b4c89039f6f9876e3185b83a7f02e7fa67ae9448e9","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"pixel-lab11-baseline-evidence","title":"进入 LAB11：保留简单基线再看条件变化","number":3},"sectionTitle":"同一组分割实验里，简单模型为何值得留下","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p36","l3-r2-p38"],"knowledgeBlock":"pixel-model-positive-result","formulas":[],"previousBlock":"b-728ba1710711069633cb474c","nextBlock":"b-ec9882427b0887ee99930f43","prerequisites":[{"id":"b-728ba1710711069633cb474c","url":"#b-728ba1710711069633cb474c","role":"section-title"},{"id":"b-8092bdd2f0e5e62669986b13","url":"#b-8092bdd2f0e5e62669986b13","role":"problem-context"}],"relatedBlocks":[{"id":"b-728ba1710711069633cb474c","url":"#b-728ba1710711069633cb474c","relation":"same-reasoning"},{"id":"b-ec9882427b0887ee99930f43","url":"#b-ec9882427b0887ee99930f43","relation":"same-reasoning"},{"id":"b-f8fd42219407d6858d415af0","url":"#b-f8fd42219407d6858d415af0","relation":"same-reasoning"},{"id":"b-c01b9fdea0abd68a152e49aa","url":"#b-c01b9fdea0abd68a152e49aa","relation":"same-reasoning"}]},{"id":"b-ec9882427b0887ee99930f43","type":"p","scope":"pixel-model-positive-result","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-ec9882427b0887ee99930f43","content":"已有 2026-09-22 本地完整运行中，原始合成图的 RGB 基线三类 mIoU=1，FCN=.99041；对输入做 5×5 模糊后，二者变为 .84996 与 .92397。原分布上简单模型胜出说明颜色已足够决定类别；模糊使边缘颜色混合，FCN 利用邻域在该干预下更稳健。两项观察共同说明为什么需要基线：它既暴露任务太容易的地方，也帮助找出复杂结构实际增加能力的条件。","sha256":"01b07d138cf5170bb6f10b5ab91b026d706baa6e021a922543b13d85cf3271ac","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"pixel-lab11-baseline-evidence","title":"进入 LAB11：保留简单基线再看条件变化","number":3},"sectionTitle":"同一组分割实验里，简单模型为何值得留下","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p36","l3-r2-p38"],"knowledgeBlock":"pixel-model-positive-result","formulas":[],"previousBlock":"b-7cb21e34c6f920a5ef64b4f0","nextBlock":"b-f8fd42219407d6858d415af0","prerequisites":[{"id":"b-728ba1710711069633cb474c","url":"#b-728ba1710711069633cb474c","role":"section-title"},{"id":"b-8092bdd2f0e5e62669986b13","url":"#b-8092bdd2f0e5e62669986b13","role":"problem-context"}],"relatedBlocks":[{"id":"b-728ba1710711069633cb474c","url":"#b-728ba1710711069633cb474c","relation":"same-reasoning"},{"id":"b-7cb21e34c6f920a5ef64b4f0","url":"#b-7cb21e34c6f920a5ef64b4f0","relation":"same-reasoning"},{"id":"b-f8fd42219407d6858d415af0","url":"#b-f8fd42219407d6858d415af0","relation":"same-reasoning"},{"id":"b-c01b9fdea0abd68a152e49aa","url":"#b-c01b9fdea0abd68a152e49aa","relation":"same-reasoning"}]},{"id":"b-f8fd42219407d6858d415af0","type":"p","scope":"pixel-model-positive-result","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-f8fd42219407d6858d415af0","content":"同一 FCN 参数，推理时仅把双线性改为最近邻，mIoU 变成 .91180。这检验训练/推理采样约定的敏感性，没有分别重训两种网络，不能当成插值方式的普遍排名。","sha256":"53b208f148eb09e2dac3c533a2268e625a9dd513bee413b9a41b925e40631934","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"pixel-lab11-baseline-evidence","title":"进入 LAB11：保留简单基线再看条件变化","number":3},"sectionTitle":"同一组分割实验里，简单模型为何值得留下","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p36","l3-r2-p38"],"knowledgeBlock":"pixel-model-positive-result","formulas":[],"previousBlock":"b-ec9882427b0887ee99930f43","nextBlock":"b-c01b9fdea0abd68a152e49aa","prerequisites":[{"id":"b-728ba1710711069633cb474c","url":"#b-728ba1710711069633cb474c","role":"section-title"},{"id":"b-8092bdd2f0e5e62669986b13","url":"#b-8092bdd2f0e5e62669986b13","role":"problem-context"}],"relatedBlocks":[{"id":"b-728ba1710711069633cb474c","url":"#b-728ba1710711069633cb474c","relation":"same-reasoning"},{"id":"b-7cb21e34c6f920a5ef64b4f0","url":"#b-7cb21e34c6f920a5ef64b4f0","relation":"same-reasoning"},{"id":"b-ec9882427b0887ee99930f43","url":"#b-ec9882427b0887ee99930f43","relation":"same-reasoning"},{"id":"b-c01b9fdea0abd68a152e49aa","url":"#b-c01b9fdea0abd68a152e49aa","relation":"same-reasoning"}]},{"id":"b-c01b9fdea0abd68a152e49aa","type":"p","scope":"pixel-model-positive-result","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-c01b9fdea0abd68a152e49aa","content":"本轮整理只复用已有记录，没有重跑。完整学习证据在 RGB→模型→logits→mask 链条；P37 的已知 mask 缩放热身则解释采样，二者都保留，作用分清。","sha256":"c31efb764a8e17b83caed97f41768f2bcebbb1a1964a23bcc8cbf6462d1b09a0","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"pixel-lab11-baseline-evidence","title":"进入 LAB11：保留简单基线再看条件变化","number":3},"sectionTitle":"同一组分割实验里，简单模型为何值得留下","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p36","l3-r2-p38"],"knowledgeBlock":"pixel-model-positive-result","formulas":[],"previousBlock":"b-f8fd42219407d6858d415af0","nextBlock":"b-41c09ebcaca31c3ba48bc0b8","prerequisites":[{"id":"b-728ba1710711069633cb474c","url":"#b-728ba1710711069633cb474c","role":"section-title"},{"id":"b-8092bdd2f0e5e62669986b13","url":"#b-8092bdd2f0e5e62669986b13","role":"problem-context"}],"relatedBlocks":[{"id":"b-728ba1710711069633cb474c","url":"#b-728ba1710711069633cb474c","relation":"same-reasoning"},{"id":"b-7cb21e34c6f920a5ef64b4f0","url":"#b-7cb21e34c6f920a5ef64b4f0","relation":"same-reasoning"},{"id":"b-ec9882427b0887ee99930f43","url":"#b-ec9882427b0887ee99930f43","relation":"same-reasoning"},{"id":"b-f8fd42219407d6858d415af0","url":"#b-f8fd42219407d6858d415af0","relation":"same-reasoning"}]},{"id":"b-41c09ebcaca31c3ba48bc0b8","type":"h2","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-41c09ebcaca31c3ba48bc0b8","content":"## 沿这个问题继续","sha256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"pixel-lab11-baseline-evidence","title":"进入 LAB11：保留简单基线再看条件变化","number":3},"sectionTitle":"进入 LAB11：保留简单基线再看条件变化","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p36","l3-r2-p38"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c01b9fdea0abd68a152e49aa","nextBlock":"b-eb316c507541bdcfeb765cd2","prerequisites":[{"id":"b-728ba1710711069633cb474c","url":"#b-728ba1710711069633cb474c","role":"section-title"},{"id":"b-8092bdd2f0e5e62669986b13","url":"#b-8092bdd2f0e5e62669986b13","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-eb316c507541bdcfeb765cd2","type":"p","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-eb316c507541bdcfeb765cd2","content":"沿 CLS08 的像素目标与指标阅读，再选 LAB11 查看已有预测和错误图。","sha256":"30703647a4a582d742b4724deb64e595de3208a7897216648bca33414cceabd0","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"pixel-lab11-baseline-evidence","title":"进入 LAB11：保留简单基线再看条件变化","number":3},"sectionTitle":"进入 LAB11：保留简单基线再看条件变化","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p36","l3-r2-p38"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-41c09ebcaca31c3ba48bc0b8","nextBlock":"b-350fd972acbb50750d3671ac","prerequisites":[{"id":"b-41c09ebcaca31c3ba48bc0b8","url":"#b-41c09ebcaca31c3ba48bc0b8","role":"section-title"},{"id":"b-8092bdd2f0e5e62669986b13","url":"#b-8092bdd2f0e5e62669986b13","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-350fd972acbb50750d3671ac","type":"h4","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p36.html#b-350fd972acbb50750d3671ac","content":"#### 按问题回查的讲解与短练习（1）","sha256":"0480869f0a66009b015a757d67aed6061011e378987f98a024f4e8bfc5386152","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"pixel-lab11-baseline-evidence","title":"进入 LAB11：保留简单基线再看条件变化","number":3},"sectionTitle":"进入 LAB11：保留简单基线再看条件变化","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p36","l3-r2-p38"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-eb316c507541bdcfeb765cd2","nextBlock":null,"prerequisites":[{"id":"b-41c09ebcaca31c3ba48bc0b8","url":"#b-41c09ebcaca31c3ba48bc0b8","role":"section-title"},{"id":"b-8092bdd2f0e5e62669986b13","url":"#b-8092bdd2f0e5e62669986b13","role":"problem-context"}],"relatedBlocks":[]}]}