{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"输出变细，不意味着信息自动变多","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html","markdown":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.md","pages":[{"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":1},{"id":"pixel-upsampling-input","title":"把粗网格放大时，先看输入有没有变","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-mask-metrics","title":"用同一张 8×8 mask 核对四种指标","unit":"怎样直接学习一张像素类别地图？","unitId":"pixel-prediction-objective","chapter":"II","number":3},{"id":"pixel-fine-coarse-fusion","title":"给预测模块接入另一份细位置信息","unit":"哪些信息真的改善了像素预测？","unitId":"pixel-detail-and-baselines","chapter":"II","number":4}],"aliases":[{"oldId":"explanation","targetId":"pixel-upsampling-input","operation":"split-explanation-range"},{"oldId":"b-2b46b28b7bc7c603c56931fc","targetId":"b-2b46b28b7bc7c603c56931fc","page":"pixel-supervision-contract","operation":"move","beforeSHA256":"f568ceb25e144397ab1a1ca1b9d741354418524c98eef10422cc026f6412eb5f","afterSHA256":"f568ceb25e144397ab1a1ca1b9d741354418524c98eef10422cc026f6412eb5f","reason":"Same source body in authored semantic page"},{"oldId":"b-88987cb4bb2bebe921e4f500","targetId":"b-88987cb4bb2bebe921e4f500","page":"pixel-supervision-contract","operation":"move","beforeSHA256":"eb0dfe6cebaefe68f519f69350ef606a2c687649403ed826851930d7434e55db","afterSHA256":"eb0dfe6cebaefe68f519f69350ef606a2c687649403ed826851930d7434e55db","reason":"Same source body in authored semantic page"},{"oldId":"b-c339465114fb80518fcd5b3b","targetId":"b-c339465114fb80518fcd5b3b","page":"pixel-supervision-contract","operation":"move","beforeSHA256":"ab380c1e0f9d1f36683516a23e6ae4f936078b69369c6e93f7772b9db9492422","afterSHA256":"ab380c1e0f9d1f36683516a23e6ae4f936078b69369c6e93f7772b9db9492422","reason":"Same source body in authored semantic page"},{"oldId":"b-e3ab168d11f56ecbc10c2b1a","targetId":"b-e3ab168d11f56ecbc10c2b1a","page":"pixel-supervision-contract","operation":"move","beforeSHA256":"a4c0dbd56af877d0b5e594f748f6f8d9603d840d6fc9a5838318be6e455ab3d5","afterSHA256":"a4c0dbd56af877d0b5e594f748f6f8d9603d840d6fc9a5838318be6e455ab3d5","reason":"Same source body in authored semantic page"},{"oldId":"b-a0101375e354e50b4a4218ff","targetId":"b-a0101375e354e50b4a4218ff","page":"pixel-supervision-contract","operation":"move","beforeSHA256":"fb67cbca43315249869971d864733783ccb53739e70b92e40c20d6581533bb82","afterSHA256":"fb67cbca43315249869971d864733783ccb53739e70b92e40c20d6581533bb82","reason":"Same source body in authored semantic page"},{"oldId":"b-145fa44a37eb1a3e82907057","targetId":"b-145fa44a37eb1a3e82907057","page":"pixel-supervision-contract","operation":"move","beforeSHA256":"75a26f3c334073d2322457d1f81001561d6c818c8f83b13b0f038ae1d77e5cc7","afterSHA256":"75a26f3c334073d2322457d1f81001561d6c818c8f83b13b0f038ae1d77e5cc7","reason":"Same source body in authored semantic page"},{"oldId":"b-e0fbe2dc10edeae75204123b","targetId":"b-e0fbe2dc10edeae75204123b","page":"pixel-upsampling-input","operation":"move","beforeSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","afterSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","reason":"Same source body in authored semantic page"},{"oldId":"b-e3594c66a906682695f39fd0","targetId":"b-e3594c66a906682695f39fd0","page":"pixel-upsampling-input","operation":"move","beforeSHA256":"dd32d09f2d9e584c55af01cc7f0149877b76480dbc3300a5ab992cace0e88994","afterSHA256":"dd32d09f2d9e584c55af01cc7f0149877b76480dbc3300a5ab992cace0e88994","reason":"Same source body in authored semantic page"},{"oldId":"b-5a5760d5266ea240fd1f1cee","targetId":"b-5a5760d5266ea240fd1f1cee","page":"pixel-upsampling-input","operation":"move","beforeSHA256":"1c12d6c2d1f8cd98252583075212849047aca363619a16470da6c06b0d620968","afterSHA256":"1c12d6c2d1f8cd98252583075212849047aca363619a16470da6c06b0d620968","reason":"Same source body in authored semantic page"},{"oldId":"b-3992d0aaa085e723ee70db0f","targetId":"b-3992d0aaa085e723ee70db0f","page":"pixel-upsampling-input","operation":"rewrite","beforeSHA256":"150325738b8dac0cd5a954c6f09f5ed1985767c7418d2812944da111f413c3c8","afterSHA256":"1f39978cfd9a3aeeb44bed7377df1ea97d4df5f1bc95da51f4ba53863e42c904","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-e4228d1d32843f9cbc2d1476","targetId":"b-e4228d1d32843f9cbc2d1476","page":"pixel-upsampling-input","operation":"move","beforeSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","afterSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","reason":"Same source body in authored semantic page"},{"oldId":"b-4f3f45578f826ac62a4a86fd","targetId":"b-4f3f45578f826ac62a4a86fd","page":"pixel-upsampling-input","operation":"move","beforeSHA256":"be54770adf55b9e0cc22b37339f79d1fffb6655def9206f88e56a0add9a193cb","afterSHA256":"be54770adf55b9e0cc22b37339f79d1fffb6655def9206f88e56a0add9a193cb","reason":"Same source body in authored semantic page"},{"oldId":"b-d5d4dc59a897340ec7eb719b","targetId":"b-d5d4dc59a897340ec7eb719b","page":"pixel-upsampling-input","operation":"move","beforeSHA256":"b3d5c753047b09c8390683f561aef1a2e4b4a5465c583a3c8fe459b563b869e5","afterSHA256":"b3d5c753047b09c8390683f561aef1a2e4b4a5465c583a3c8fe459b563b869e5","reason":"Same source body in authored semantic page"},{"oldId":"b-d1e9a1da1195a327fb9763fc","targetId":"b-d1e9a1da1195a327fb9763fc","page":"pixel-upsampling-input","operation":"move","beforeSHA256":"8fe1536177bc1aa8ea763fd58962d54b821ae21880d5a1f7a6614f0020052ba0","afterSHA256":"8fe1536177bc1aa8ea763fd58962d54b821ae21880d5a1f7a6614f0020052ba0","reason":"Same source body in authored semantic page"},{"oldId":"b-bb2a57d4019178d2a019b597","targetId":"b-bb2a57d4019178d2a019b597","page":"pixel-upsampling-input","operation":"move","beforeSHA256":"98f880b9989e4811358611984c824fe8a03547b6e6226e2e32ba809afbe392cd","afterSHA256":"98f880b9989e4811358611984c824fe8a03547b6e6226e2e32ba809afbe392cd","reason":"Same source body in authored semantic page"},{"oldId":"b-39f432a9b255505cdfffac39","targetId":"b-39f432a9b255505cdfffac39","page":"pixel-upsampling-input","operation":"move","beforeSHA256":"2420c851e3ef42fad47e3d15f6708e98b3eb9ec0e577cd57cceed5ae1f67e66b","afterSHA256":"2420c851e3ef42fad47e3d15f6708e98b3eb9ec0e577cd57cceed5ae1f67e66b","reason":"Same source body in authored semantic page"},{"oldId":"b-f1e189607a33a65d289d9a7e","targetId":"b-f1e189607a33a65d289d9a7e","page":"pixel-upsampling-input","operation":"move","beforeSHA256":"3fc38d54a0a14d9962a04f11fa9a5d95de99c74dab5dcb41f793d1e877e06038","afterSHA256":"3fc38d54a0a14d9962a04f11fa9a5d95de99c74dab5dcb41f793d1e877e06038","reason":"Same source body in authored semantic page"},{"oldId":"b-a5818ac29eb0b8361e4e6dbd","targetId":"b-a5818ac29eb0b8361e4e6dbd","page":"pixel-upsampling-input","operation":"move","beforeSHA256":"e613ad3c3ec0234d8e09f245f2bf20847d3c4df0c977b65ebbca8daca7fd3254","afterSHA256":"e613ad3c3ec0234d8e09f245f2bf20847d3c4df0c977b65ebbca8daca7fd3254","reason":"Same source body in authored semantic page"},{"oldId":"b-0d8dfab4dc32054d1a3a524c","targetId":"b-0d8dfab4dc32054d1a3a524c","page":"pixel-upsampling-input","operation":"move","beforeSHA256":"c821102f0e664e51749eeb3bd95c8f703895eb7c0853e39785e9e4abd2dfaf3e","afterSHA256":"c821102f0e664e51749eeb3bd95c8f703895eb7c0853e39785e9e4abd2dfaf3e","reason":"Same source body in authored semantic page"},{"oldId":"b-3ddfd84fae67797e09b8def8","targetId":"b-3ddfd84fae67797e09b8def8","page":"pixel-upsampling-input","operation":"move","beforeSHA256":"c3108fe3e1d7ac9f3f2070a59aa06b2308dfccbaab555b613c5575c1e4aea69d","afterSHA256":"c3108fe3e1d7ac9f3f2070a59aa06b2308dfccbaab555b613c5575c1e4aea69d","reason":"Same source body in authored semantic page"},{"oldId":"b-8505fdc5bbf3ecc1447e19ce","targetId":"b-8505fdc5bbf3ecc1447e19ce","page":"pixel-mask-metrics","operation":"move","beforeSHA256":"7146794e11b242cb4d3808858a7758adbedd82ccb76a7f1f42213f509bd43fea","afterSHA256":"7146794e11b242cb4d3808858a7758adbedd82ccb76a7f1f42213f509bd43fea","reason":"Same source body in authored semantic page"},{"oldId":"b-2f8b9a016651df679c416fd1","targetId":"b-2f8b9a016651df679c416fd1","page":"pixel-mask-metrics","operation":"move","beforeSHA256":"2abef8836bc8264b308f99a51ee96a2155fb5eff95c9c70505a9677db7c5cec2","afterSHA256":"2abef8836bc8264b308f99a51ee96a2155fb5eff95c9c70505a9677db7c5cec2","reason":"Same source body in authored semantic page"},{"oldId":"b-caddbcabf58ceba0e8482b19","targetId":"b-caddbcabf58ceba0e8482b19","page":"pixel-mask-metrics","operation":"rewrite","beforeSHA256":"1c9dca60e949866fe3d2c1e845c73d1a487da3b37a8b825f9336a144a1d3c544","afterSHA256":"749503cc42b305dbf612f9950e053a8bd470fd67b6b03c07fab171fa53f4a3cd","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-7cd4610154fcb94ffecce082","targetId":"b-7cd4610154fcb94ffecce082","page":"pixel-mask-metrics","operation":"move","beforeSHA256":"7901720f8cf1d01174de7937f30f30409cc15448d8c9d52d834f9f5d39f015da","afterSHA256":"7901720f8cf1d01174de7937f30f30409cc15448d8c9d52d834f9f5d39f015da","reason":"Same source body in authored semantic page"},{"oldId":"b-9ed2921d2a9c29dcd64b727a","targetId":"b-9ed2921d2a9c29dcd64b727a","page":"pixel-mask-metrics","operation":"move","beforeSHA256":"c10e7af18ab5805c53068a7fef8219df87144cfc145a5fffab03b20d3f2c9304","afterSHA256":"c10e7af18ab5805c53068a7fef8219df87144cfc145a5fffab03b20d3f2c9304","reason":"Same source body in authored semantic page"},{"oldId":"b-be155c2dfd0af6ab268b5229","targetId":"b-be155c2dfd0af6ab268b5229","page":"pixel-fine-coarse-fusion","operation":"move","beforeSHA256":"bac9bc0bfc7367f973a9f54acb560f8b718933f35a4984dc618ac3811f018cf4","afterSHA256":"bac9bc0bfc7367f973a9f54acb560f8b718933f35a4984dc618ac3811f018cf4","reason":"Same source body in authored semantic page"},{"oldId":"b-4c56690f1811ea8283eb6b03","targetId":"b-4c56690f1811ea8283eb6b03","page":"pixel-fine-coarse-fusion","operation":"move","beforeSHA256":"43da12e890070d8c515227c5ebc5c2a02bda955d0bbc9814ce7274df52f7e15e","afterSHA256":"43da12e890070d8c515227c5ebc5c2a02bda955d0bbc9814ce7274df52f7e15e","reason":"Same source body in authored semantic page"},{"oldId":"b-f485a023fe09143fb88d214d","targetId":"b-f485a023fe09143fb88d214d","page":"pixel-fine-coarse-fusion","operation":"move","beforeSHA256":"251c7a6709bc4329ed718f766ff1a82e10bc886b6fd57056d1c9d7df1e69b6c3","afterSHA256":"251c7a6709bc4329ed718f766ff1a82e10bc886b6fd57056d1c9d7df1e69b6c3","reason":"Same source body in authored semantic page"},{"oldId":"b-ca4c2b59cae624e51fbb81b6","targetId":"b-ca4c2b59cae624e51fbb81b6","page":"pixel-fine-coarse-fusion","operation":"move","beforeSHA256":"47f4c1dbdddeb4c55f8650fb141a134e01362d41b4b320d3383c7aba567ca1d0","afterSHA256":"47f4c1dbdddeb4c55f8650fb141a134e01362d41b4b320d3383c7aba567ca1d0","reason":"Same source body in authored semantic page"},{"oldId":"b-99e7be46b041eeedae51e291","targetId":"b-99e7be46b041eeedae51e291","page":"pixel-fine-coarse-fusion","operation":"move","beforeSHA256":"245a6c3894b4038ad3b6211589ae7ffe3d009b1254334cf573991a14eff28c1e","afterSHA256":"245a6c3894b4038ad3b6211589ae7ffe3d009b1254334cf573991a14eff28c1e","reason":"Same source body in authored semantic page"},{"oldId":"b-4709a470988dc4f634826af0","targetId":"b-4709a470988dc4f634826af0","page":"pixel-fine-coarse-fusion","operation":"move","beforeSHA256":"8c27dec41c2374067fc317c3cce617d7fabe5478ed174bacef8698d34965b0ee","afterSHA256":"8c27dec41c2374067fc317c3cce617d7fabe5478ed174bacef8698d34965b0ee","reason":"Same source body in authored semantic page"},{"oldId":"b-83d7ab9c965c674653e746aa","targetId":"b-83d7ab9c965c674653e746aa","page":"pixel-fine-coarse-fusion","operation":"move","beforeSHA256":"f16be0424762a44344338696f06ef266c4571d15367ae2cbbd28fac98ac7e9cb","afterSHA256":"f16be0424762a44344338696f06ef266c4571d15367ae2cbbd28fac98ac7e9cb","reason":"Same source body in authored semantic page"},{"oldId":"b-c8dbe15ad4f45bf8a8ec0613","targetId":"b-c8dbe15ad4f45bf8a8ec0613","page":"pixel-fine-coarse-fusion","operation":"move","beforeSHA256":"b25685f04cd1b885f3a5999e6a912460c7424021cdc5b2190584242086a82681","afterSHA256":"b25685f04cd1b885f3a5999e6a912460c7424021cdc5b2190584242086a82681","reason":"Same source body in authored semantic page"},{"oldId":"b-36da76f1d6f78556637fbb31","targetId":"b-36da76f1d6f78556637fbb31","page":"pixel-fine-coarse-fusion","operation":"move","beforeSHA256":"ece621ebd76199d1445fe08bdf7683dd75d5a9567d80ac60173df99d8b074096","afterSHA256":"ece621ebd76199d1445fe08bdf7683dd75d5a9567d80ac60173df99d8b074096","reason":"Same source body in authored semantic page"},{"oldId":"b-c3c213454cef3dfddcb145db","targetId":"b-c3c213454cef3dfddcb145db","page":"pixel-fine-coarse-fusion","operation":"move","beforeSHA256":"8550d64dd381526fafe330b62e1c6d55c7dbd8ca9fd86378c1c8f572e1f871c0","afterSHA256":"8550d64dd381526fafe330b62e1c6d55c7dbd8ca9fd86378c1c8f572e1f871c0","reason":"Same source body in authored semantic page"},{"oldId":"b-269624641512c7dfe373f358","targetId":"b-269624641512c7dfe373f358","page":"pixel-fine-coarse-fusion","operation":"move","beforeSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","afterSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","reason":"Same source body in authored semantic page"},{"oldId":"b-924875301446934c9dd9b830","targetId":"b-924875301446934c9dd9b830","page":"pixel-fine-coarse-fusion","operation":"move","beforeSHA256":"30703647a4a582d742b4724deb64e595de3208a7897216648bca33414cceabd0","afterSHA256":"30703647a4a582d742b4724deb64e595de3208a7897216648bca33414cceabd0","reason":"Same source body in authored semantic page"},{"oldId":"pixel-supervision-contract-title","targetId":"b-5916dd2033f37d05abba8bee","operation":"retain-source-identity"},{"oldId":"pixel-supervision-contract-bridge","targetId":"b-c0ca621bb9c799aeca3b48f6","operation":"retain-source-identity"},{"oldId":"pixel-upsampling-input-title","targetId":"b-cf01de2a9d65be426abf9be1","operation":"retain-source-identity"},{"oldId":"pixel-mask-metrics-title","targetId":"b-6eb9f2d4bc5c9d9c985aa69f","operation":"retain-source-identity"},{"oldId":"pixel-mask-metrics-bridge","targetId":"b-afafbee42662ec3417c6fe43","operation":"retain-source-identity"},{"oldId":"pixel-fine-coarse-fusion-title","targetId":"b-475c57cf03bfc961118a7d8e","operation":"retain-source-identity"},{"oldId":"pixel-fine-coarse-fusion-bridge","targetId":"b-bf4c5603fde9596b25ae8a90","operation":"retain-source-identity"}],"blocks":[{"id":"b-5916dd2033f37d05abba8bee","type":"h2","scope":"pixel-supervision-contract","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-5916dd2033f37d05abba8bee","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":1},"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-c0ca621bb9c799aeca3b48f6","prerequisites":[{"id":"b-c0ca621bb9c799aeca3b48f6","url":"#b-c0ca621bb9c799aeca3b48f6","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-c0ca621bb9c799aeca3b48f6","type":"p","scope":"pixel-supervision-contract","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-c0ca621bb9c799aeca3b48f6","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":1},"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-5916dd2033f37d05abba8bee","nextBlock":"b-2b46b28b7bc7c603c56931fc","prerequisites":[{"id":"b-5916dd2033f37d05abba8bee","url":"#b-5916dd2033f37d05abba8bee","role":"section-title"}],"relatedBlocks":[]},{"id":"b-2b46b28b7bc7c603c56931fc","type":"h3","scope":"pixel-cross-entropy","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-2b46b28b7bc7c603c56931fc","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":1},"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-c0ca621bb9c799aeca3b48f6","nextBlock":"b-88987cb4bb2bebe921e4f500","prerequisites":[{"id":"b-5916dd2033f37d05abba8bee","url":"#b-5916dd2033f37d05abba8bee","role":"section-title"},{"id":"b-c0ca621bb9c799aeca3b48f6","url":"#b-c0ca621bb9c799aeca3b48f6","role":"problem-context"}],"relatedBlocks":[{"id":"b-88987cb4bb2bebe921e4f500","url":"#b-88987cb4bb2bebe921e4f500","relation":"same-reasoning"},{"id":"b-c339465114fb80518fcd5b3b","url":"#b-c339465114fb80518fcd5b3b","relation":"same-reasoning"},{"id":"b-e3ab168d11f56ecbc10c2b1a","url":"#b-e3ab168d11f56ecbc10c2b1a","relation":"same-reasoning"},{"id":"b-a0101375e354e50b4a4218ff","url":"#b-a0101375e354e50b4a4218ff","relation":"same-reasoning"},{"id":"b-145fa44a37eb1a3e82907057","url":"#b-145fa44a37eb1a3e82907057","relation":"same-reasoning"}]},{"id":"b-88987cb4bb2bebe921e4f500","type":"p","scope":"pixel-cross-entropy","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-88987cb4bb2bebe921e4f500","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":1},"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-2b46b28b7bc7c603c56931fc","nextBlock":"b-c339465114fb80518fcd5b3b","prerequisites":[{"id":"b-2b46b28b7bc7c603c56931fc","url":"#b-2b46b28b7bc7c603c56931fc","role":"section-title"},{"id":"b-c0ca621bb9c799aeca3b48f6","url":"#b-c0ca621bb9c799aeca3b48f6","role":"problem-context"}],"relatedBlocks":[{"id":"b-2b46b28b7bc7c603c56931fc","url":"#b-2b46b28b7bc7c603c56931fc","relation":"same-reasoning"},{"id":"b-c339465114fb80518fcd5b3b","url":"#b-c339465114fb80518fcd5b3b","relation":"same-reasoning"},{"id":"b-e3ab168d11f56ecbc10c2b1a","url":"#b-e3ab168d11f56ecbc10c2b1a","relation":"same-reasoning"},{"id":"b-a0101375e354e50b4a4218ff","url":"#b-a0101375e354e50b4a4218ff","relation":"same-reasoning"},{"id":"b-145fa44a37eb1a3e82907057","url":"#b-145fa44a37eb1a3e82907057","relation":"same-reasoning"}]},{"id":"b-c339465114fb80518fcd5b3b","type":"p","scope":"pixel-cross-entropy","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-c339465114fb80518fcd5b3b","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":1},"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-88987cb4bb2bebe921e4f500","nextBlock":"b-e3ab168d11f56ecbc10c2b1a","prerequisites":[{"id":"b-2b46b28b7bc7c603c56931fc","url":"#b-2b46b28b7bc7c603c56931fc","role":"section-title"},{"id":"b-c0ca621bb9c799aeca3b48f6","url":"#b-c0ca621bb9c799aeca3b48f6","role":"problem-context"}],"relatedBlocks":[{"id":"b-2b46b28b7bc7c603c56931fc","url":"#b-2b46b28b7bc7c603c56931fc","relation":"same-reasoning"},{"id":"b-88987cb4bb2bebe921e4f500","url":"#b-88987cb4bb2bebe921e4f500","relation":"same-reasoning"},{"id":"b-e3ab168d11f56ecbc10c2b1a","url":"#b-e3ab168d11f56ecbc10c2b1a","relation":"same-reasoning"},{"id":"b-a0101375e354e50b4a4218ff","url":"#b-a0101375e354e50b4a4218ff","relation":"same-reasoning"},{"id":"b-145fa44a37eb1a3e82907057","url":"#b-145fa44a37eb1a3e82907057","relation":"same-reasoning"}]},{"id":"b-e3ab168d11f56ecbc10c2b1a","type":"p","scope":"pixel-cross-entropy","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-e3ab168d11f56ecbc10c2b1a","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":1},"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-c339465114fb80518fcd5b3b","nextBlock":"b-a0101375e354e50b4a4218ff","prerequisites":[{"id":"b-2b46b28b7bc7c603c56931fc","url":"#b-2b46b28b7bc7c603c56931fc","role":"section-title"},{"id":"b-c0ca621bb9c799aeca3b48f6","url":"#b-c0ca621bb9c799aeca3b48f6","role":"problem-context"}],"relatedBlocks":[{"id":"b-2b46b28b7bc7c603c56931fc","url":"#b-2b46b28b7bc7c603c56931fc","relation":"same-reasoning"},{"id":"b-88987cb4bb2bebe921e4f500","url":"#b-88987cb4bb2bebe921e4f500","relation":"same-reasoning"},{"id":"b-c339465114fb80518fcd5b3b","url":"#b-c339465114fb80518fcd5b3b","relation":"same-reasoning"},{"id":"b-a0101375e354e50b4a4218ff","url":"#b-a0101375e354e50b4a4218ff","relation":"same-reasoning"},{"id":"b-145fa44a37eb1a3e82907057","url":"#b-145fa44a37eb1a3e82907057","relation":"same-reasoning"}]},{"id":"b-a0101375e354e50b4a4218ff","type":"p","scope":"pixel-cross-entropy","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-a0101375e354e50b4a4218ff","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":1},"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-e3ab168d11f56ecbc10c2b1a","nextBlock":"b-145fa44a37eb1a3e82907057","prerequisites":[{"id":"b-2b46b28b7bc7c603c56931fc","url":"#b-2b46b28b7bc7c603c56931fc","role":"section-title"},{"id":"b-c0ca621bb9c799aeca3b48f6","url":"#b-c0ca621bb9c799aeca3b48f6","role":"problem-context"}],"relatedBlocks":[{"id":"b-2b46b28b7bc7c603c56931fc","url":"#b-2b46b28b7bc7c603c56931fc","relation":"same-reasoning"},{"id":"b-88987cb4bb2bebe921e4f500","url":"#b-88987cb4bb2bebe921e4f500","relation":"same-reasoning"},{"id":"b-c339465114fb80518fcd5b3b","url":"#b-c339465114fb80518fcd5b3b","relation":"same-reasoning"},{"id":"b-e3ab168d11f56ecbc10c2b1a","url":"#b-e3ab168d11f56ecbc10c2b1a","relation":"same-reasoning"},{"id":"b-145fa44a37eb1a3e82907057","url":"#b-145fa44a37eb1a3e82907057","relation":"same-reasoning"}]},{"id":"b-145fa44a37eb1a3e82907057","type":"p","scope":"pixel-cross-entropy","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-145fa44a37eb1a3e82907057","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":1},"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-a0101375e354e50b4a4218ff","nextBlock":null,"prerequisites":[{"id":"b-2b46b28b7bc7c603c56931fc","url":"#b-2b46b28b7bc7c603c56931fc","role":"section-title"},{"id":"b-c0ca621bb9c799aeca3b48f6","url":"#b-c0ca621bb9c799aeca3b48f6","role":"problem-context"}],"relatedBlocks":[{"id":"b-2b46b28b7bc7c603c56931fc","url":"#b-2b46b28b7bc7c603c56931fc","relation":"same-reasoning"},{"id":"b-88987cb4bb2bebe921e4f500","url":"#b-88987cb4bb2bebe921e4f500","relation":"same-reasoning"},{"id":"b-c339465114fb80518fcd5b3b","url":"#b-c339465114fb80518fcd5b3b","relation":"same-reasoning"},{"id":"b-e3ab168d11f56ecbc10c2b1a","url":"#b-e3ab168d11f56ecbc10c2b1a","relation":"same-reasoning"},{"id":"b-a0101375e354e50b4a4218ff","url":"#b-a0101375e354e50b4a4218ff","relation":"same-reasoning"}]},{"id":"b-cf01de2a9d65be426abf9be1","type":"h2","scope":"pixel-upsampling-input","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-cf01de2a9d65be426abf9be1","content":"## 把粗网格放大时，先看输入有没有变","sha256":"8d1f73f0a9994f3acf69f816a1b1675e1d3ff90f97fd81375823bc5712735cc2","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-e0fbe2dc10edeae75204123b","prerequisites":[{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-e0fbe2dc10edeae75204123b","type":"h2","scope":"explanation","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-e0fbe2dc10edeae75204123b","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-cf01de2a9d65be426abf9be1","nextBlock":"b-e3594c66a906682695f39fd0","prerequisites":[{"id":"b-cf01de2a9d65be426abf9be1","url":"#b-cf01de2a9d65be426abf9be1","role":"section-title"},{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-e3594c66a906682695f39fd0","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-e3594c66a906682695f39fd0","content":"# P37 · 输出变细，不意味着信息自动变多","sha256":"dd32d09f2d9e584c55af01cc7f0149877b76480dbc3300a5ab992cace0e88994","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e0fbe2dc10edeae75204123b","nextBlock":"b-5a5760d5266ea240fd1f1cee","prerequisites":[{"id":"b-e0fbe2dc10edeae75204123b","url":"#b-e0fbe2dc10edeae75204123b","role":"section-title"},{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-5a5760d5266ea240fd1f1cee","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-5a5760d5266ea240fd1f1cee","content":"比较增加输出格子与增加输入证据。最近邻和双线性只规定怎样利用现有数值；离散标签与连续 logits 的插值又不同。8×8 算例清楚给出前景 IoU、背景 IoU、mIoU 和像素准确率，避免同一个小数换四个名字。","sha256":"1c12d6c2d1f8cd98252583075212849047aca363619a16470da6c06b0d620968","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e3594c66a906682695f39fd0","nextBlock":"b-3992d0aaa085e723ee70db0f","prerequisites":[{"id":"b-e3594c66a906682695f39fd0","url":"#b-e3594c66a906682695f39fd0","role":"section-title"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-3992d0aaa085e723ee70db0f","type":"figure","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-3992d0aaa085e723ee70db0f","content":"图内标签与数值：Lecture 03 课堂图解 P37；已有 4 × 4 数值；最近邻 · 16 × 16；双线性 · 16 × 16；唯一的数值来源；复制邻近值；过渡更平滑；两种放大都只读原来的 16 个值，没有重新读取图像。；二值数值场的插值算例；不是视觉模型预测。双线性按输出像素中心映射，边界夹取。\n\n同一个 4×4 二值数值场放大到 16×16，最近邻和双线性并排；两者的计算输入完全相同。\n\n- 已有 4 × 4 数值\n- 最近邻 · 16 × 16\n- 双线性 · 16 × 16\n- 唯一的数值来源\n- 复制邻近值\n- 过渡更平滑\n- 两种放大都只读原来的 16 个值，没有重新读取图像。\n- 二值数值场的插值算例；不是视觉模型预测。双线性按输出像素中心映射，边界夹取。同一个 4×4 二值数值场放大到 16×16，最近邻和双线性并排；两者的计算输入完全相同。 [查看对应图解状态](/#/l3-r2-p37/0)[查看完整图解](/course/figures/l3-r2-p37-0.html)","sha256":"1f39978cfd9a3aeeb44bed7377df1ea97d4df5f1bc95da51f4ba53863e42c904","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"同一个 4×4 二值数值场放大到 16×16，最近邻和双线性并排；两者的计算输入完全相同。 查看对应图解状态","labels":["已有 4 × 4 数值","最近邻 · 16 × 16","双线性 · 16 × 16","唯一的数值来源","复制邻近值","过渡更平滑","两种放大都只读原来的 16 个值，没有重新读取图像。","二值数值场的插值算例；不是视觉模型预测。双线性按输出像素中心映射，边界夹取。"],"description":"同一个 4×4 二值数值场放大到 16×16，最近邻和双线性并排；两者的计算输入完全相同。已有 4 × 4 数值最近邻 · 16 × 16双线性 · 16 × 16唯一的数值来源复制邻近值过渡更平滑两种放大都只读原来的 16 个值，没有重新读取图像。二值数值场的插值算例；不是视觉模型预测。双线性按输出像素中心映射，边界夹取。","evidenceIdentity":"同一个 4×4 二值数值场放大到 16×16，最近邻和双线性并排；两者的计算输入完全相同。 查看对应图解状态","images":[]},"previousBlock":"b-5a5760d5266ea240fd1f1cee","nextBlock":"b-e4228d1d32843f9cbc2d1476","prerequisites":[{"id":"b-e3594c66a906682695f39fd0","url":"#b-e3594c66a906682695f39fd0","role":"section-title"},{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-e4228d1d32843f9cbc2d1476","type":"h2","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-e4228d1d32843f9cbc2d1476","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3992d0aaa085e723ee70db0f","nextBlock":"b-4f3f45578f826ac62a4a86fd","prerequisites":[{"id":"b-e3594c66a906682695f39fd0","url":"#b-e3594c66a906682695f39fd0","role":"section-title"},{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-4f3f45578f826ac62a4a86fd","type":"p","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-4f3f45578f826ac62a4a86fd","content":"把一个粗网格放大很容易。最近邻复制数值，双线性利用邻近位置作平滑过渡；它们可以给出更多输出位置，但没有额外读取原图。\n如果深层表示中仍保存着相关差异，合适的解码和上采样能帮助利用它；如果两种不同边界已经被编码成完全一样的表示，单纯放大不会知道原来是哪一种。\n所以像素预测不仅需要“把图片尺寸变回来”，还需要检查表示中保留的空间信息、解码时的读取和训练监督。不要把转置卷积叫成一般意义上的逆卷积，也不要把画面更平滑等同于边界更正确。下面看怎样把不同层次的信息结合起来。","sha256":"be54770adf55b9e0cc22b37339f79d1fffb6655def9206f88e56a0add9a193cb","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e4228d1d32843f9cbc2d1476","nextBlock":"b-92fd781056bf0cde686cfadd","prerequisites":[{"id":"b-e4228d1d32843f9cbc2d1476","url":"#b-e4228d1d32843f9cbc2d1476","role":"section-title"},{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-92fd781056bf0cde686cfadd","type":"h4","scope":"pixel-upsampling-input","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-92fd781056bf0cde686cfadd","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-4f3f45578f826ac62a4a86fd","nextBlock":"b-d5d4dc59a897340ec7eb719b","prerequisites":[{"id":"b-e4228d1d32843f9cbc2d1476","url":"#b-e4228d1d32843f9cbc2d1476","role":"section-title"},{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-d5d4dc59a897340ec7eb719b","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-d5d4dc59a897340ec7eb719b","content":"同一 4×4 二值数值场，最近邻 floor(outindex×4/16)；双线性按 (outindex+.5)×4/16−.5 映射并对边界夹取。可比较插值，不称为从图像预测 mask 的训练实验。","sha256":"b3d5c753047b09c8390683f561aef1a2e4b4a5465c583a3c8fe459b563b869e5","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-92fd781056bf0cde686cfadd","nextBlock":"b-9aecd8726cbe89fdcc7d8b11","prerequisites":[{"id":"b-92fd781056bf0cde686cfadd","url":"#b-92fd781056bf0cde686cfadd","role":"section-title"},{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-9aecd8726cbe89fdcc7d8b11","type":"h4","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-9aecd8726cbe89fdcc7d8b11","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d5d4dc59a897340ec7eb719b","nextBlock":"b-d1e9a1da1195a327fb9763fc","prerequisites":[{"id":"b-92fd781056bf0cde686cfadd","url":"#b-92fd781056bf0cde686cfadd","role":"section-title"},{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-d1e9a1da1195a327fb9763fc","type":"ul","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-d1e9a1da1195a327fb9763fc","content":"- [FCN 是独立的密集预测路线｜一起算一个完整例子](/supplement.html?return=l3-r2-p37/0#/l3-fcn-dense-example/0) 本页说明\n- [上采样不会自动恢复信息｜问题从哪里出现](/supplement.html?return=l3-r2-p37/0#/l3-upsampling-problem/0) L03补充阅读\n- [上采样不会自动恢复信息｜核心改动怎样工作](/supplement.html?return=l3-r2-p37/0#/l3-upsampling-mechanism/0) L03补充阅读\n- [上采样不会自动恢复信息｜一起算一个完整例子](/supplement.html?return=l3-r2-p37/0#/l3-upsampling-example/0) L03补充阅读\n- [上采样不会自动恢复信息｜怎样检查这项解释](/supplement.html?return=l3-r2-p37/0#/l3-upsampling-evidence/0) L03补充阅读\n- [上采样不会自动恢复信息｜什么结论还不能推出](/supplement.html?return=l3-r2-p37/0#/l3-upsampling-boundary/0) L03补充阅读","sha256":"8fe1536177bc1aa8ea763fd58962d54b821ae21880d5a1f7a6614f0020052ba0","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9aecd8726cbe89fdcc7d8b11","nextBlock":"b-bb2a57d4019178d2a019b597","prerequisites":[{"id":"b-9aecd8726cbe89fdcc7d8b11","url":"#b-9aecd8726cbe89fdcc7d8b11","role":"section-title"},{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-bb2a57d4019178d2a019b597","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-bb2a57d4019178d2a019b597","content":"- [FCN 是独立的密集预测路线｜一起算一个完整例子](/supplement.html?return=l3-r2-p37/0#/l3-fcn-dense-example/0) 本页说明","sha256":"98f880b9989e4811358611984c824fe8a03547b6e6226e2e32ba809afbe392cd","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d1e9a1da1195a327fb9763fc","nextBlock":"b-39f432a9b255505cdfffac39","prerequisites":[{"id":"b-9aecd8726cbe89fdcc7d8b11","url":"#b-9aecd8726cbe89fdcc7d8b11","role":"section-title"},{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-39f432a9b255505cdfffac39","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-39f432a9b255505cdfffac39","content":"- [上采样不会自动恢复信息｜问题从哪里出现](/supplement.html?return=l3-r2-p37/0#/l3-upsampling-problem/0) L03补充阅读","sha256":"2420c851e3ef42fad47e3d15f6708e98b3eb9ec0e577cd57cceed5ae1f67e66b","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-bb2a57d4019178d2a019b597","nextBlock":"b-f1e189607a33a65d289d9a7e","prerequisites":[{"id":"b-9aecd8726cbe89fdcc7d8b11","url":"#b-9aecd8726cbe89fdcc7d8b11","role":"section-title"},{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-f1e189607a33a65d289d9a7e","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-f1e189607a33a65d289d9a7e","content":"- [上采样不会自动恢复信息｜核心改动怎样工作](/supplement.html?return=l3-r2-p37/0#/l3-upsampling-mechanism/0) L03补充阅读","sha256":"3fc38d54a0a14d9962a04f11fa9a5d95de99c74dab5dcb41f793d1e877e06038","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-39f432a9b255505cdfffac39","nextBlock":"b-a5818ac29eb0b8361e4e6dbd","prerequisites":[{"id":"b-9aecd8726cbe89fdcc7d8b11","url":"#b-9aecd8726cbe89fdcc7d8b11","role":"section-title"},{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-a5818ac29eb0b8361e4e6dbd","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-a5818ac29eb0b8361e4e6dbd","content":"- [上采样不会自动恢复信息｜一起算一个完整例子](/supplement.html?return=l3-r2-p37/0#/l3-upsampling-example/0) L03补充阅读","sha256":"e613ad3c3ec0234d8e09f245f2bf20847d3c4df0c977b65ebbca8daca7fd3254","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f1e189607a33a65d289d9a7e","nextBlock":"b-0d8dfab4dc32054d1a3a524c","prerequisites":[{"id":"b-9aecd8726cbe89fdcc7d8b11","url":"#b-9aecd8726cbe89fdcc7d8b11","role":"section-title"},{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-0d8dfab4dc32054d1a3a524c","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-0d8dfab4dc32054d1a3a524c","content":"- [上采样不会自动恢复信息｜怎样检查这项解释](/supplement.html?return=l3-r2-p37/0#/l3-upsampling-evidence/0) L03补充阅读","sha256":"c821102f0e664e51749eeb3bd95c8f703895eb7c0853e39785e9e4abd2dfaf3e","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-a5818ac29eb0b8361e4e6dbd","nextBlock":"b-3ddfd84fae67797e09b8def8","prerequisites":[{"id":"b-9aecd8726cbe89fdcc7d8b11","url":"#b-9aecd8726cbe89fdcc7d8b11","role":"section-title"},{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-3ddfd84fae67797e09b8def8","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-3ddfd84fae67797e09b8def8","content":"- [上采样不会自动恢复信息｜什么结论还不能推出](/supplement.html?return=l3-r2-p37/0#/l3-upsampling-boundary/0) L03补充阅读","sha256":"c3108fe3e1d7ac9f3f2070a59aa06b2308dfccbaab555b613c5575c1e4aea69d","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-upsampling-input","title":"把粗网格放大时，先看输入有没有变","number":2},"sectionTitle":"把粗网格放大时，先看输入有没有变","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-0d8dfab4dc32054d1a3a524c","nextBlock":null,"prerequisites":[{"id":"b-9aecd8726cbe89fdcc7d8b11","url":"#b-9aecd8726cbe89fdcc7d8b11","role":"section-title"},{"id":"b-5a5760d5266ea240fd1f1cee","url":"#b-5a5760d5266ea240fd1f1cee","role":"problem-context"}],"relatedBlocks":[{"id":"b-3992d0aaa085e723ee70db0f","url":"#b-3992d0aaa085e723ee70db0f","relation":"figure-and-explanation"}]},{"id":"b-6eb9f2d4bc5c9d9c985aa69f","type":"h2","scope":"pixel-mask-metrics","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-6eb9f2d4bc5c9d9c985aa69f","content":"## 用同一张 8×8 mask 核对四种指标","sha256":"832b48dbbbd2b064ff329afb7938a95af99fe7ee5d8c601ea262b2c1fad0d5d6","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-mask-metrics","title":"用同一张 8×8 mask 核对四种指标","number":3},"sectionTitle":"用同一张 8×8 mask 核对四种指标","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-afafbee42662ec3417c6fe43","prerequisites":[{"id":"b-afafbee42662ec3417c6fe43","url":"#b-afafbee42662ec3417c6fe43","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-afafbee42662ec3417c6fe43","type":"p","scope":"pixel-mask-metrics","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-afafbee42662ec3417c6fe43","content":"已知 mask 的缩放练习隔离采样；相同错误要按明确口径计数。","sha256":"0a0ab9e1b5b9e9116cea0b559ff033935d7e692479179fa716ff626b87bc25c4","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-mask-metrics","title":"用同一张 8×8 mask 核对四种指标","number":3},"sectionTitle":"用同一张 8×8 mask 核对四种指标","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-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6eb9f2d4bc5c9d9c985aa69f","nextBlock":"b-8505fdc5bbf3ecc1447e19ce","prerequisites":[{"id":"b-6eb9f2d4bc5c9d9c985aa69f","url":"#b-6eb9f2d4bc5c9d9c985aa69f","role":"section-title"}],"relatedBlocks":[]},{"id":"b-8505fdc5bbf3ecc1447e19ce","type":"h3","scope":"mask-iou-worked-example","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-8505fdc5bbf3ecc1447e19ce","content":"### 0.667 是前景 IoU，0.75 才是本例两类 mIoU","sha256":"7146794e11b242cb4d3808858a7758adbedd82ccb76a7f1f42213f509bd43fea","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-mask-metrics","title":"用同一张 8×8 mask 核对四种指标","number":3},"sectionTitle":"0.667 是前景 IoU，0.75 才是本例两类 mIoU","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-p37"],"knowledgeBlock":"mask-iou-worked-example","formulas":[],"previousBlock":"b-afafbee42662ec3417c6fe43","nextBlock":"b-2f8b9a016651df679c416fd1","prerequisites":[{"id":"b-6eb9f2d4bc5c9d9c985aa69f","url":"#b-6eb9f2d4bc5c9d9c985aa69f","role":"section-title"},{"id":"b-afafbee42662ec3417c6fe43","url":"#b-afafbee42662ec3417c6fe43","role":"problem-context"}],"relatedBlocks":[{"id":"b-2f8b9a016651df679c416fd1","url":"#b-2f8b9a016651df679c416fd1","relation":"same-reasoning"},{"id":"b-caddbcabf58ceba0e8482b19","url":"#b-caddbcabf58ceba0e8482b19","relation":"same-reasoning"},{"id":"b-7cd4610154fcb94ffecce082","url":"#b-7cd4610154fcb94ffecce082","relation":"same-reasoning"},{"id":"b-9ed2921d2a9c29dcd64b727a","url":"#b-9ed2921d2a9c29dcd64b727a","relation":"same-reasoning"}]},{"id":"b-2f8b9a016651df679c416fd1","type":"p","scope":"mask-iou-worked-example","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-2f8b9a016651df679c416fd1","content":"已知 8×8 真值 mask：前景 16、背景 48。平均降到 4×4，再最近邻还原并以 ≥.5 阈值化后，预测前景 24，其中 16 个正确、8 个误把背景当成前景；没有漏掉真前景。因此 TP=16、FP=8、FN=0、TN=40。","sha256":"2abef8836bc8264b308f99a51ee96a2155fb5eff95c9c70505a9677db7c5cec2","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-mask-metrics","title":"用同一张 8×8 mask 核对四种指标","number":3},"sectionTitle":"0.667 是前景 IoU，0.75 才是本例两类 mIoU","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-p37"],"knowledgeBlock":"mask-iou-worked-example","formulas":[],"previousBlock":"b-8505fdc5bbf3ecc1447e19ce","nextBlock":"b-caddbcabf58ceba0e8482b19","prerequisites":[{"id":"b-8505fdc5bbf3ecc1447e19ce","url":"#b-8505fdc5bbf3ecc1447e19ce","role":"section-title"},{"id":"b-afafbee42662ec3417c6fe43","url":"#b-afafbee42662ec3417c6fe43","role":"problem-context"}],"relatedBlocks":[{"id":"b-8505fdc5bbf3ecc1447e19ce","url":"#b-8505fdc5bbf3ecc1447e19ce","relation":"same-reasoning"},{"id":"b-caddbcabf58ceba0e8482b19","url":"#b-caddbcabf58ceba0e8482b19","relation":"same-reasoning"},{"id":"b-7cd4610154fcb94ffecce082","url":"#b-7cd4610154fcb94ffecce082","relation":"same-reasoning"},{"id":"b-9ed2921d2a9c29dcd64b727a","url":"#b-9ed2921d2a9c29dcd64b727a","relation":"same-reasoning"}]},{"id":"b-caddbcabf58ceba0e8482b19","type":"table","scope":"mask-iou-worked-example","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-caddbcabf58ceba0e8482b19","content":"| 口径 | 分子与分母 | 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 | .75 |\n| 像素准确率 | (16+40)/64 | .875 |","sha256":"749503cc42b305dbf612f9950e053a8bd470fd67b6b03c07fab171fa53f4a3cd","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-mask-metrics","title":"用同一张 8×8 mask 核对四种指标","number":3},"sectionTitle":"0.667 是前景 IoU，0.75 才是本例两类 mIoU","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-p37"],"knowledgeBlock":"mask-iou-worked-example","formulas":[],"table":{"headers":["口径","分子与分母","结果"],"rows":[["口径","分子与分母","结果"],["前景 IoU","16/(16+8+0)","2/3"],["背景 IoU","40/(40+0+8)","5/6"],["两类 mIoU，含背景","(2/3+5/6)/2",".75"],["像素准确率","(16+40)/64",".875"]]},"previousBlock":"b-2f8b9a016651df679c416fd1","nextBlock":"b-7cd4610154fcb94ffecce082","prerequisites":[{"id":"b-8505fdc5bbf3ecc1447e19ce","url":"#b-8505fdc5bbf3ecc1447e19ce","role":"section-title"},{"id":"b-afafbee42662ec3417c6fe43","url":"#b-afafbee42662ec3417c6fe43","role":"problem-context"}],"relatedBlocks":[{"id":"b-8505fdc5bbf3ecc1447e19ce","url":"#b-8505fdc5bbf3ecc1447e19ce","relation":"same-reasoning"},{"id":"b-2f8b9a016651df679c416fd1","url":"#b-2f8b9a016651df679c416fd1","relation":"same-reasoning"},{"id":"b-7cd4610154fcb94ffecce082","url":"#b-7cd4610154fcb94ffecce082","relation":"same-reasoning"},{"id":"b-9ed2921d2a9c29dcd64b727a","url":"#b-9ed2921d2a9c29dcd64b727a","relation":"same-reasoning"}]},{"id":"b-7cd4610154fcb94ffecce082","type":"p","scope":"mask-iou-worked-example","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-7cd4610154fcb94ffecce082","content":"这个过程的输入就是已知 mask，因此它隔离的是采样与阈值，不是看 RGB 图像做分割。当前 LAB11 已将名称修正，并保留最近邻与双线性两个分支。双线性分支的两类 mIoU 为 .77564，前景 IoU 仍为 2/3；平滑改变的是错误的位置和类别组成，不是自动保证每项指标都变好。","sha256":"7901720f8cf1d01174de7937f30f30409cc15448d8c9d52d834f9f5d39f015da","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-mask-metrics","title":"用同一张 8×8 mask 核对四种指标","number":3},"sectionTitle":"0.667 是前景 IoU，0.75 才是本例两类 mIoU","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-p37"],"knowledgeBlock":"mask-iou-worked-example","formulas":[],"previousBlock":"b-caddbcabf58ceba0e8482b19","nextBlock":"b-9ed2921d2a9c29dcd64b727a","prerequisites":[{"id":"b-8505fdc5bbf3ecc1447e19ce","url":"#b-8505fdc5bbf3ecc1447e19ce","role":"section-title"},{"id":"b-afafbee42662ec3417c6fe43","url":"#b-afafbee42662ec3417c6fe43","role":"problem-context"}],"relatedBlocks":[{"id":"b-8505fdc5bbf3ecc1447e19ce","url":"#b-8505fdc5bbf3ecc1447e19ce","relation":"same-reasoning"},{"id":"b-2f8b9a016651df679c416fd1","url":"#b-2f8b9a016651df679c416fd1","relation":"same-reasoning"},{"id":"b-caddbcabf58ceba0e8482b19","url":"#b-caddbcabf58ceba0e8482b19","relation":"same-reasoning"},{"id":"b-9ed2921d2a9c29dcd64b727a","url":"#b-9ed2921d2a9c29dcd64b727a","relation":"same-reasoning"}]},{"id":"b-9ed2921d2a9c29dcd64b727a","type":"p","scope":"mask-iou-worked-example","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-9ed2921d2a9c29dcd64b727a","content":"读一项 mIoU 时至少应知道平均哪些类、是否含背景、空并集怎样处理，以及先累计整份混淆矩阵还是逐图计算再平均。这些口径会改变数字；具体口径比多显示小数位更重要。","sha256":"c10e7af18ab5805c53068a7fef8219df87144cfc145a5fffab03b20d3f2c9304","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-mask-metrics","title":"用同一张 8×8 mask 核对四种指标","number":3},"sectionTitle":"0.667 是前景 IoU，0.75 才是本例两类 mIoU","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-p37"],"knowledgeBlock":"mask-iou-worked-example","formulas":[],"previousBlock":"b-7cd4610154fcb94ffecce082","nextBlock":null,"prerequisites":[{"id":"b-8505fdc5bbf3ecc1447e19ce","url":"#b-8505fdc5bbf3ecc1447e19ce","role":"section-title"},{"id":"b-afafbee42662ec3417c6fe43","url":"#b-afafbee42662ec3417c6fe43","role":"problem-context"}],"relatedBlocks":[{"id":"b-8505fdc5bbf3ecc1447e19ce","url":"#b-8505fdc5bbf3ecc1447e19ce","relation":"same-reasoning"},{"id":"b-2f8b9a016651df679c416fd1","url":"#b-2f8b9a016651df679c416fd1","relation":"same-reasoning"},{"id":"b-caddbcabf58ceba0e8482b19","url":"#b-caddbcabf58ceba0e8482b19","relation":"same-reasoning"},{"id":"b-7cd4610154fcb94ffecce082","url":"#b-7cd4610154fcb94ffecce082","relation":"same-reasoning"}]},{"id":"b-475c57cf03bfc961118a7d8e","type":"h2","scope":"pixel-fine-coarse-fusion","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-475c57cf03bfc961118a7d8e","content":"## 给预测模块接入另一份细位置信息","sha256":"7c84d52c1f8cbf85583aff0dcbf51067c5749d36c08e5406f1cd748009b1626d","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-fine-coarse-fusion","title":"给预测模块接入另一份细位置信息","number":4},"sectionTitle":"给预测模块接入另一份细位置信息","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p38","l3-r2-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-bf4c5603fde9596b25ae8a90","prerequisites":[{"id":"b-bf4c5603fde9596b25ae8a90","url":"#b-bf4c5603fde9596b25ae8a90","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-bf4c5603fde9596b25ae8a90","type":"p","scope":"pixel-fine-coarse-fusion","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-bf4c5603fde9596b25ae8a90","content":"上采样读取现有数值，跨层连接则把另一份特征交给预测模块。","sha256":"c8a41bd6e3cc137742d70a38c19f1af043cbed3efa432d3c5b1f0a7854728166","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-fine-coarse-fusion","title":"给预测模块接入另一份细位置信息","number":4},"sectionTitle":"给预测模块接入另一份细位置信息","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p38","l3-r2-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-475c57cf03bfc961118a7d8e","nextBlock":"b-be155c2dfd0af6ab268b5229","prerequisites":[{"id":"b-475c57cf03bfc961118a7d8e","url":"#b-475c57cf03bfc961118a7d8e","role":"section-title"}],"relatedBlocks":[]},{"id":"b-be155c2dfd0af6ab268b5229","type":"h3","scope":"upsampling-versus-new-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-be155c2dfd0af6ab268b5229","content":"### 上采样改变读数位置，跨层连接提供额外证据","sha256":"bac9bc0bfc7367f973a9f54acb560f8b718933f35a4984dc618ac3811f018cf4","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-fine-coarse-fusion","title":"给预测模块接入另一份细位置信息","number":4},"sectionTitle":"上采样改变读数位置，跨层连接提供额外证据","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p38","l3-r2-p37"],"knowledgeBlock":"upsampling-versus-new-evidence","formulas":[],"previousBlock":"b-bf4c5603fde9596b25ae8a90","nextBlock":"b-4c56690f1811ea8283eb6b03","prerequisites":[{"id":"b-475c57cf03bfc961118a7d8e","url":"#b-475c57cf03bfc961118a7d8e","role":"section-title"},{"id":"b-bf4c5603fde9596b25ae8a90","url":"#b-bf4c5603fde9596b25ae8a90","role":"problem-context"}],"relatedBlocks":[{"id":"b-4c56690f1811ea8283eb6b03","url":"#b-4c56690f1811ea8283eb6b03","relation":"same-reasoning"},{"id":"b-f485a023fe09143fb88d214d","url":"#b-f485a023fe09143fb88d214d","relation":"same-reasoning"},{"id":"b-ca4c2b59cae624e51fbb81b6","url":"#b-ca4c2b59cae624e51fbb81b6","relation":"same-reasoning"},{"id":"b-99e7be46b041eeedae51e291","url":"#b-99e7be46b041eeedae51e291","relation":"same-reasoning"},{"id":"b-4709a470988dc4f634826af0","url":"#b-4709a470988dc4f634826af0","relation":"same-reasoning"}]},{"id":"b-4c56690f1811ea8283eb6b03","type":"p","scope":"upsampling-versus-new-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-4c56690f1811ea8283eb6b03","content":"从 12×12 分数图恢复到 48×48，可以用最近邻复制，也可以用双线性按邻域加权。输出格子多了，是为了在原图坐标上作决定；运算本身没有再次看原图。若细边界的信息已被编码成相同数值，单独放大同一张图不能判断它原来在哪。","sha256":"43da12e890070d8c515227c5ebc5c2a02bda955d0bbc9814ce7274df52f7e15e","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-fine-coarse-fusion","title":"给预测模块接入另一份细位置信息","number":4},"sectionTitle":"上采样改变读数位置，跨层连接提供额外证据","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p38","l3-r2-p37"],"knowledgeBlock":"upsampling-versus-new-evidence","formulas":[],"previousBlock":"b-be155c2dfd0af6ab268b5229","nextBlock":"b-f485a023fe09143fb88d214d","prerequisites":[{"id":"b-be155c2dfd0af6ab268b5229","url":"#b-be155c2dfd0af6ab268b5229","role":"section-title"},{"id":"b-bf4c5603fde9596b25ae8a90","url":"#b-bf4c5603fde9596b25ae8a90","role":"problem-context"}],"relatedBlocks":[{"id":"b-be155c2dfd0af6ab268b5229","url":"#b-be155c2dfd0af6ab268b5229","relation":"same-reasoning"},{"id":"b-f485a023fe09143fb88d214d","url":"#b-f485a023fe09143fb88d214d","relation":"same-reasoning"},{"id":"b-ca4c2b59cae624e51fbb81b6","url":"#b-ca4c2b59cae624e51fbb81b6","relation":"same-reasoning"},{"id":"b-99e7be46b041eeedae51e291","url":"#b-99e7be46b041eeedae51e291","relation":"same-reasoning"},{"id":"b-4709a470988dc4f634826af0","url":"#b-4709a470988dc4f634826af0","relation":"same-reasoning"}]},{"id":"b-f485a023fe09143fb88d214d","type":"p","scope":"upsampling-versus-new-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-f485a023fe09143fb88d214d","content":"但低分辨率并不意味着一切细节必丢。通道、相位与邻域响应可能保留可用线索，训练出的解码器可以利用它们；真正的判断是“差别有没有进入 F”，而不只是 F 的宽高。因此 P37 既不许诺放大恢复真实细节，也不否定学习上采样的价值。","sha256":"251c7a6709bc4329ed718f766ff1a82e10bc886b6fd57056d1c9d7df1e69b6c3","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-fine-coarse-fusion","title":"给预测模块接入另一份细位置信息","number":4},"sectionTitle":"上采样改变读数位置，跨层连接提供额外证据","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p38","l3-r2-p37"],"knowledgeBlock":"upsampling-versus-new-evidence","formulas":[],"previousBlock":"b-4c56690f1811ea8283eb6b03","nextBlock":"b-ca4c2b59cae624e51fbb81b6","prerequisites":[{"id":"b-be155c2dfd0af6ab268b5229","url":"#b-be155c2dfd0af6ab268b5229","role":"section-title"},{"id":"b-bf4c5603fde9596b25ae8a90","url":"#b-bf4c5603fde9596b25ae8a90","role":"problem-context"}],"relatedBlocks":[{"id":"b-be155c2dfd0af6ab268b5229","url":"#b-be155c2dfd0af6ab268b5229","relation":"same-reasoning"},{"id":"b-4c56690f1811ea8283eb6b03","url":"#b-4c56690f1811ea8283eb6b03","relation":"same-reasoning"},{"id":"b-ca4c2b59cae624e51fbb81b6","url":"#b-ca4c2b59cae624e51fbb81b6","relation":"same-reasoning"},{"id":"b-99e7be46b041eeedae51e291","url":"#b-99e7be46b041eeedae51e291","relation":"same-reasoning"},{"id":"b-4709a470988dc4f634826af0","url":"#b-4709a470988dc4f634826af0","relation":"same-reasoning"}]},{"id":"b-ca4c2b59cae624e51fbb81b6","type":"p","scope":"upsampling-versus-new-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-ca4c2b59cae624e51fbb81b6","content":"跨层连接做了一件不同的事：把较细层的特征另行送进预测模块。它提供的不是同一粗特征的复制，而是另一份尚可能保留位置细节的输入。深层上下文可帮助识别这是杯子，较细响应可帮助定位杯柄与窄缝；二者仍需通过任务训练学习怎样组合。","sha256":"47f4c1dbdddeb4c55f8650fb141a134e01362d41b4b320d3383c7aba567ca1d0","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-fine-coarse-fusion","title":"给预测模块接入另一份细位置信息","number":4},"sectionTitle":"上采样改变读数位置，跨层连接提供额外证据","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p38","l3-r2-p37"],"knowledgeBlock":"upsampling-versus-new-evidence","formulas":[],"previousBlock":"b-f485a023fe09143fb88d214d","nextBlock":"b-99e7be46b041eeedae51e291","prerequisites":[{"id":"b-be155c2dfd0af6ab268b5229","url":"#b-be155c2dfd0af6ab268b5229","role":"section-title"},{"id":"b-bf4c5603fde9596b25ae8a90","url":"#b-bf4c5603fde9596b25ae8a90","role":"problem-context"}],"relatedBlocks":[{"id":"b-be155c2dfd0af6ab268b5229","url":"#b-be155c2dfd0af6ab268b5229","relation":"same-reasoning"},{"id":"b-4c56690f1811ea8283eb6b03","url":"#b-4c56690f1811ea8283eb6b03","relation":"same-reasoning"},{"id":"b-f485a023fe09143fb88d214d","url":"#b-f485a023fe09143fb88d214d","relation":"same-reasoning"},{"id":"b-99e7be46b041eeedae51e291","url":"#b-99e7be46b041eeedae51e291","relation":"same-reasoning"},{"id":"b-4709a470988dc4f634826af0","url":"#b-4709a470988dc4f634826af0","relation":"same-reasoning"}]},{"id":"b-99e7be46b041eeedae51e291","type":"p","scope":"upsampling-versus-new-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-99e7be46b041eeedae51e291","content":"转置卷积是可学习的线性运算形式，不是一般意义上可还原任意输入的“逆卷积”。是否保留细节，应通过同源输入、目标、预测与错误图观察。","sha256":"245a6c3894b4038ad3b6211589ae7ffe3d009b1254334cf573991a14eff28c1e","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-fine-coarse-fusion","title":"给预测模块接入另一份细位置信息","number":4},"sectionTitle":"上采样改变读数位置，跨层连接提供额外证据","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p38","l3-r2-p37"],"knowledgeBlock":"upsampling-versus-new-evidence","formulas":[],"previousBlock":"b-ca4c2b59cae624e51fbb81b6","nextBlock":"b-4709a470988dc4f634826af0","prerequisites":[{"id":"b-be155c2dfd0af6ab268b5229","url":"#b-be155c2dfd0af6ab268b5229","role":"section-title"},{"id":"b-bf4c5603fde9596b25ae8a90","url":"#b-bf4c5603fde9596b25ae8a90","role":"problem-context"}],"relatedBlocks":[{"id":"b-be155c2dfd0af6ab268b5229","url":"#b-be155c2dfd0af6ab268b5229","relation":"same-reasoning"},{"id":"b-4c56690f1811ea8283eb6b03","url":"#b-4c56690f1811ea8283eb6b03","relation":"same-reasoning"},{"id":"b-f485a023fe09143fb88d214d","url":"#b-f485a023fe09143fb88d214d","relation":"same-reasoning"},{"id":"b-ca4c2b59cae624e51fbb81b6","url":"#b-ca4c2b59cae624e51fbb81b6","relation":"same-reasoning"},{"id":"b-4709a470988dc4f634826af0","url":"#b-4709a470988dc4f634826af0","relation":"same-reasoning"}]},{"id":"b-4709a470988dc4f634826af0","type":"p","scope":"upsampling-versus-new-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-4709a470988dc4f634826af0","content":"P38 的杯柄照片只是原图参考：细、粗特征先做坐标对齐，再由预测模块学习组合，交付像素类别分数或掩码，而不是还原那张 RGB 照片。图中的特征格为结构示意，轮廓为教学标注；是否减少空洞和边界错误，要另看模型预测与目标的比较。","sha256":"8c27dec41c2374067fc317c3cce617d7fabe5478ed174bacef8698d34965b0ee","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-fine-coarse-fusion","title":"给预测模块接入另一份细位置信息","number":4},"sectionTitle":"上采样改变读数位置，跨层连接提供额外证据","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p38","l3-r2-p37"],"knowledgeBlock":"upsampling-versus-new-evidence","formulas":[],"previousBlock":"b-99e7be46b041eeedae51e291","nextBlock":"b-83d7ab9c965c674653e746aa","prerequisites":[{"id":"b-be155c2dfd0af6ab268b5229","url":"#b-be155c2dfd0af6ab268b5229","role":"section-title"},{"id":"b-bf4c5603fde9596b25ae8a90","url":"#b-bf4c5603fde9596b25ae8a90","role":"problem-context"}],"relatedBlocks":[{"id":"b-be155c2dfd0af6ab268b5229","url":"#b-be155c2dfd0af6ab268b5229","relation":"same-reasoning"},{"id":"b-4c56690f1811ea8283eb6b03","url":"#b-4c56690f1811ea8283eb6b03","relation":"same-reasoning"},{"id":"b-f485a023fe09143fb88d214d","url":"#b-f485a023fe09143fb88d214d","relation":"same-reasoning"},{"id":"b-ca4c2b59cae624e51fbb81b6","url":"#b-ca4c2b59cae624e51fbb81b6","relation":"same-reasoning"},{"id":"b-99e7be46b041eeedae51e291","url":"#b-99e7be46b041eeedae51e291","relation":"same-reasoning"}]},{"id":"b-83d7ab9c965c674653e746aa","type":"h3","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-83d7ab9c965c674653e746aa","content":"### 沿着图解追问","sha256":"f16be0424762a44344338696f06ef266c4571d15367ae2cbbd28fac98ac7e9cb","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-fine-coarse-fusion","title":"给预测模块接入另一份细位置信息","number":4},"sectionTitle":"给预测模块接入另一份细位置信息","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p38","l3-r2-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-4709a470988dc4f634826af0","nextBlock":"b-c8dbe15ad4f45bf8a8ec0613","prerequisites":[{"id":"b-be155c2dfd0af6ab268b5229","url":"#b-be155c2dfd0af6ab268b5229","role":"section-title"},{"id":"b-bf4c5603fde9596b25ae8a90","url":"#b-bf4c5603fde9596b25ae8a90","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-c8dbe15ad4f45bf8a8ec0613","type":"p","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-c8dbe15ad4f45bf8a8ec0613","content":"以下是备课用的模拟师生对话，不是课堂实录。","sha256":"b25685f04cd1b885f3a5999e6a912460c7424021cdc5b2190584242086a82681","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-fine-coarse-fusion","title":"给预测模块接入另一份细位置信息","number":4},"sectionTitle":"给预测模块接入另一份细位置信息","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p38","l3-r2-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-83d7ab9c965c674653e746aa","nextBlock":"b-36da76f1d6f78556637fbb31","prerequisites":[{"id":"b-83d7ab9c965c674653e746aa","url":"#b-83d7ab9c965c674653e746aa","role":"section-title"},{"id":"b-bf4c5603fde9596b25ae8a90","url":"#b-bf4c5603fde9596b25ae8a90","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-36da76f1d6f78556637fbb31","type":"p","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-36da76f1d6f78556637fbb31","content":"**学生：**双线性更平滑，是不是边界就更准？","sha256":"ece621ebd76199d1445fe08bdf7683dd75d5a9567d80ac60173df99d8b074096","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-fine-coarse-fusion","title":"给预测模块接入另一份细位置信息","number":4},"sectionTitle":"给预测模块接入另一份细位置信息","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p38","l3-r2-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c8dbe15ad4f45bf8a8ec0613","nextBlock":"b-c3c213454cef3dfddcb145db","prerequisites":[{"id":"b-83d7ab9c965c674653e746aa","url":"#b-83d7ab9c965c674653e746aa","role":"section-title"},{"id":"b-bf4c5603fde9596b25ae8a90","url":"#b-bf4c5603fde9596b25ae8a90","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-c3c213454cef3dfddcb145db","type":"p","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-c3c213454cef3dfddcb145db","content":"**教师：**平滑只描述输出变化。是否更准要对照目标轮廓与同一指标；当前已知 mask 热身中，前景 IoU 并没有因双线性变好。","sha256":"8550d64dd381526fafe330b62e1c6d55c7dbd8ca9fd86378c1c8f572e1f871c0","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-fine-coarse-fusion","title":"给预测模块接入另一份细位置信息","number":4},"sectionTitle":"给预测模块接入另一份细位置信息","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p38","l3-r2-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-36da76f1d6f78556637fbb31","nextBlock":"b-269624641512c7dfe373f358","prerequisites":[{"id":"b-83d7ab9c965c674653e746aa","url":"#b-83d7ab9c965c674653e746aa","role":"section-title"},{"id":"b-bf4c5603fde9596b25ae8a90","url":"#b-bf4c5603fde9596b25ae8a90","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-269624641512c7dfe373f358","type":"h2","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-269624641512c7dfe373f358","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-fine-coarse-fusion","title":"给预测模块接入另一份细位置信息","number":4},"sectionTitle":"给预测模块接入另一份细位置信息","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p38","l3-r2-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c3c213454cef3dfddcb145db","nextBlock":"b-924875301446934c9dd9b830","prerequisites":[{"id":"b-83d7ab9c965c674653e746aa","url":"#b-83d7ab9c965c674653e746aa","role":"section-title"},{"id":"b-bf4c5603fde9596b25ae8a90","url":"#b-bf4c5603fde9596b25ae8a90","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-924875301446934c9dd9b830","type":"p","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p37.html#b-924875301446934c9dd9b830","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-fine-coarse-fusion","title":"给预测模块接入另一份细位置信息","number":4},"sectionTitle":"给预测模块接入另一份细位置信息","currentQuestion":"细支路、简单基线和输入干预分别能证明什么？","sourceSlides":["l3-r2-p38","l3-r2-p37"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-269624641512c7dfe373f358","nextBlock":null,"prerequisites":[{"id":"b-269624641512c7dfe373f358","url":"#b-269624641512c7dfe373f358","role":"section-title"},{"id":"b-bf4c5603fde9596b25ae8a90","url":"#b-bf4c5603fde9596b25ae8a90","role":"problem-context"}],"relatedBlocks":[]}]}