{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/blocks/mask-branch-target.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"一个 RoI 的三项输出，为什么不是三个类别标签","url":"https://codingai-lec03.pages.dev/course/blocks/mask-branch-target.html","markdown":"https://codingai-lec03.pages.dev/course/blocks/mask-branch-target.md","pages":[{"id":"reading-opening","title":"一个 RoI 的三项输出，为什么不是三个类别标签","unit":"怎样让每一份轮廓属于一个明确对象？","unitId":"instance-mask-identity","chapter":"II","number":1}],"aliases":[{"oldId":"b-eb5e44e634975dbec12ce39c","targetId":"b-eb5e44e634975dbec12ce39c","page":"reading-opening","operation":"move","beforeSHA256":"f995a64667b58a6e75d76ccdcfb8ee46ce24ebd902b694c26f8795b29b42aeb2","afterSHA256":"f995a64667b58a6e75d76ccdcfb8ee46ce24ebd902b694c26f8795b29b42aeb2","reason":"Same source body in authored semantic page"},{"oldId":"b-d445c7d76675b3f20a9c822b","targetId":"b-d445c7d76675b3f20a9c822b","page":"reading-opening","operation":"move","beforeSHA256":"1e822f009dd59e8a958ba8e5d9d7d033f7c6f4b6d131abd012857f2401cb5f5c","afterSHA256":"1e822f009dd59e8a958ba8e5d9d7d033f7c6f4b6d131abd012857f2401cb5f5c","reason":"Same source body in authored semantic page"},{"oldId":"b-4d68abb6bb44057cc1330f85","targetId":"b-4d68abb6bb44057cc1330f85","page":"reading-opening","operation":"move","beforeSHA256":"0b511a4330a11c16ba46f599835b5c88e40c9d998df27f4bf90d4c02cb2264ca","afterSHA256":"0b511a4330a11c16ba46f599835b5c88e40c9d998df27f4bf90d4c02cb2264ca","reason":"Same source body in authored semantic page"},{"oldId":"b-bebd86441907982cf0ee6f6c","targetId":"b-bebd86441907982cf0ee6f6c","page":"reading-opening","operation":"move","beforeSHA256":"157aeac42b5c38f069caaedc9a6700d329e9f40f7dc6c00ff43a490b1a1c1683","afterSHA256":"157aeac42b5c38f069caaedc9a6700d329e9f40f7dc6c00ff43a490b1a1c1683","reason":"Same source body in authored semantic page"},{"oldId":"b-3a2b8abed2e4b761b5096283","targetId":"b-3a2b8abed2e4b761b5096283","page":"reading-opening","operation":"move","beforeSHA256":"04214673903fd768071a053ae57e5f39eda5e720d70baaf627515e96a7369fb7","afterSHA256":"04214673903fd768071a053ae57e5f39eda5e720d70baaf627515e96a7369fb7","reason":"Same source body in authored semantic page"},{"oldId":"b-ca2129f787710fafd4f9d866","targetId":"b-ca2129f787710fafd4f9d866","page":"reading-opening","operation":"move","beforeSHA256":"ff964f198c4004d329d895e4061f11e6b2f004b82f8f73aa6eb186ea34262bd3","afterSHA256":"ff964f198c4004d329d895e4061f11e6b2f004b82f8f73aa6eb186ea34262bd3","reason":"Same source body in authored semantic page"},{"oldId":"b-830fe91eccf61d41fb31db19","targetId":"b-830fe91eccf61d41fb31db19","page":"reading-opening","operation":"move","beforeSHA256":"5cea2ec3cb0e60da6a03638a2746272cc387521d0c8b748c2e2c06412a9e1640","afterSHA256":"5cea2ec3cb0e60da6a03638a2746272cc387521d0c8b748c2e2c06412a9e1640","reason":"Same source body in authored semantic page"},{"oldId":"b-cc5c9bd76476d98ad9b677ec","targetId":"b-cc5c9bd76476d98ad9b677ec","page":"reading-opening","operation":"rewrite","beforeSHA256":"2780e7b5ef004a7be319ed8a23480552b05933e5ba4c75f6cd95bc2b2fe3f214","afterSHA256":"33b3dd7a7fca0aa326b276734d6d324ea0ef699894580ba29b7682a78029df04","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-0feadac72e948ffe131963c7","targetId":"b-0feadac72e948ffe131963c7","page":"reading-opening","operation":"move","beforeSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","afterSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","reason":"Same source body in authored semantic page"}],"blocks":[{"id":"b-eb5e44e634975dbec12ce39c","type":"h1","scope":"mask-branch-target","url":"https://codingai-lec03.pages.dev/course/blocks/mask-branch-target.html#b-eb5e44e634975dbec12ce39c","content":"# 一个 RoI 的三项输出，为什么不是三个类别标签","sha256":"f995a64667b58a6e75d76ccdcfb8ee46ce24ebd902b694c26f8795b29b42aeb2","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"instance-mask-identity","title":"怎样让每一份轮廓属于一个明确对象？"},"page":{"id":"reading-opening","title":"一个 RoI 的三项输出，为什么不是三个类别标签","number":1},"sectionTitle":"一个 RoI 的三项输出，为什么不是三个类别标签","currentQuestion":"一个 RoI 的三项输出，为什么不是三个类别标签","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-d445c7d76675b3f20a9c822b","prerequisites":[{"id":"b-d445c7d76675b3f20a9c822b","url":"#b-d445c7d76675b3f20a9c822b","role":"problem-context"}],"relatedBlocks":[{"id":"b-cc5c9bd76476d98ad9b677ec","url":"#b-cc5c9bd76476d98ad9b677ec","relation":"figure-and-explanation"}]},{"id":"b-d445c7d76675b3f20a9c822b","type":"p","scope":"mask-branch-target","url":"https://codingai-lec03.pages.dev/course/blocks/mask-branch-target.html#b-d445c7d76675b3f20a9c822b","content":"一个 RoI 的类别、框与局部 mask 有不同输出形状和监督。mask 分支需要二维布局，并以当前实例为目标，框中另一个同类对象不该被顺便选中。P41–43 接着检验区域读出与目标坐标是否真的一致。","sha256":"1e822f009dd59e8a958ba8e5d9d7d033f7c6f4b6d131abd012857f2401cb5f5c","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"instance-mask-identity","title":"怎样让每一份轮廓属于一个明确对象？"},"page":{"id":"reading-opening","title":"一个 RoI 的三项输出，为什么不是三个类别标签","number":1},"sectionTitle":"一个 RoI 的三项输出，为什么不是三个类别标签","currentQuestion":"一个 RoI 的三项输出，为什么不是三个类别标签","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-eb5e44e634975dbec12ce39c","nextBlock":"b-4d68abb6bb44057cc1330f85","prerequisites":[{"id":"b-eb5e44e634975dbec12ce39c","url":"#b-eb5e44e634975dbec12ce39c","role":"section-title"}],"relatedBlocks":[{"id":"b-cc5c9bd76476d98ad9b677ec","url":"#b-cc5c9bd76476d98ad9b677ec","relation":"figure-and-explanation"}]},{"id":"b-4d68abb6bb44057cc1330f85","type":"p","scope":"mask-branch-target","url":"https://codingai-lec03.pages.dev/course/blocks/mask-branch-target.html#b-4d68abb6bb44057cc1330f85","content":"同一 RoI 可以给出类别、四个框偏移和一个二维掩码。前两项可以汇总为短向量；mask 却需要保留“区域左上、右下分别是什么”的空间关系。把区域特征完全摊成不带布局的统计，会给逐像素对应增加负担。","sha256":"0b511a4330a11c16ba46f599835b5c88e40c9d998df27f4bf90d4c02cb2264ca","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"instance-mask-identity","title":"怎样让每一份轮廓属于一个明确对象？"},"page":{"id":"reading-opening","title":"一个 RoI 的三项输出，为什么不是三个类别标签","number":1},"sectionTitle":"一个 RoI 的三项输出，为什么不是三个类别标签","currentQuestion":"一个 RoI 的三项输出，为什么不是三个类别标签","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d445c7d76675b3f20a9c822b","nextBlock":"b-bebd86441907982cf0ee6f6c","prerequisites":[{"id":"b-eb5e44e634975dbec12ce39c","url":"#b-eb5e44e634975dbec12ce39c","role":"section-title"},{"id":"b-d445c7d76675b3f20a9c822b","url":"#b-d445c7d76675b3f20a9c822b","role":"problem-context"}],"relatedBlocks":[{"id":"b-cc5c9bd76476d98ad9b677ec","url":"#b-cc5c9bd76476d98ad9b677ec","relation":"figure-and-explanation"}]},{"id":"b-bebd86441907982cf0ee6f6c","type":"p","scope":"mask-branch-target","url":"https://codingai-lec03.pages.dev/course/blocks/mask-branch-target.html#b-bebd86441907982cf0ee6f6c","content":"画面中原图局部只帮助我们指认对象。真正进入区域任务头的是从共享特征 F 读取的区域特征；其空间格子保存相对位置，每格还可以有多个通道。后两页检验这一读取怎样保持连续坐标对应。","sha256":"157aeac42b5c38f069caaedc9a6700d329e9f40f7dc6c00ff43a490b1a1c1683","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"instance-mask-identity","title":"怎样让每一份轮廓属于一个明确对象？"},"page":{"id":"reading-opening","title":"一个 RoI 的三项输出，为什么不是三个类别标签","number":1},"sectionTitle":"一个 RoI 的三项输出，为什么不是三个类别标签","currentQuestion":"一个 RoI 的三项输出，为什么不是三个类别标签","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-4d68abb6bb44057cc1330f85","nextBlock":"b-3a2b8abed2e4b761b5096283","prerequisites":[{"id":"b-eb5e44e634975dbec12ce39c","url":"#b-eb5e44e634975dbec12ce39c","role":"section-title"},{"id":"b-d445c7d76675b3f20a9c822b","url":"#b-d445c7d76675b3f20a9c822b","role":"problem-context"}],"relatedBlocks":[{"id":"b-cc5c9bd76476d98ad9b677ec","url":"#b-cc5c9bd76476d98ad9b677ec","relation":"figure-and-explanation"}]},{"id":"b-3a2b8abed2e4b761b5096283","type":"p","scope":"mask-branch-target","url":"https://codingai-lec03.pages.dev/course/blocks/mask-branch-target.html#b-3a2b8abed2e4b761b5096283","content":"Mask R-CNN 的掩码分支在 RoI 内预测固定大小的二值图。原论文的默认实现对每个类别各预测一张 sigmoid mask，训练时只计算真值类别对应的掩码损失；区域分类分支负责决定类别。因此掩码像素回答的是“是否属于本实例”，不是让所有对象类别在每个 RoI 像素上再次竞争。","sha256":"04214673903fd768071a053ae57e5f39eda5e720d70baaf627515e96a7369fb7","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"instance-mask-identity","title":"怎样让每一份轮廓属于一个明确对象？"},"page":{"id":"reading-opening","title":"一个 RoI 的三项输出，为什么不是三个类别标签","number":1},"sectionTitle":"一个 RoI 的三项输出，为什么不是三个类别标签","currentQuestion":"一个 RoI 的三项输出，为什么不是三个类别标签","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-bebd86441907982cf0ee6f6c","nextBlock":"b-ca2129f787710fafd4f9d866","prerequisites":[{"id":"b-eb5e44e634975dbec12ce39c","url":"#b-eb5e44e634975dbec12ce39c","role":"section-title"},{"id":"b-d445c7d76675b3f20a9c822b","url":"#b-d445c7d76675b3f20a9c822b","role":"problem-context"}],"relatedBlocks":[{"id":"b-cc5c9bd76476d98ad9b677ec","url":"#b-cc5c9bd76476d98ad9b677ec","relation":"figure-and-explanation"}]},{"id":"b-ca2129f787710fafd4f9d866","type":"p","scope":"mask-branch-target","url":"https://codingai-lec03.pages.dev/course/blocks/mask-branch-target.html#b-ca2129f787710fafd4f9d866","content":"用杯子的例子理解：RoI 已对应杯一，框里即使还露出杯二的一部分，杯一的目标 mask 仍只标杯一。两个区域可以都被分类为杯子，分别交付各自的二值轮廓。这种分工让实例身份与像素预测接在一起。","sha256":"ff964f198c4004d329d895e4061f11e6b2f004b82f8f73aa6eb186ea34262bd3","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"instance-mask-identity","title":"怎样让每一份轮廓属于一个明确对象？"},"page":{"id":"reading-opening","title":"一个 RoI 的三项输出，为什么不是三个类别标签","number":1},"sectionTitle":"一个 RoI 的三项输出，为什么不是三个类别标签","currentQuestion":"一个 RoI 的三项输出，为什么不是三个类别标签","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3a2b8abed2e4b761b5096283","nextBlock":"b-830fe91eccf61d41fb31db19","prerequisites":[{"id":"b-eb5e44e634975dbec12ce39c","url":"#b-eb5e44e634975dbec12ce39c","role":"section-title"},{"id":"b-d445c7d76675b3f20a9c822b","url":"#b-d445c7d76675b3f20a9c822b","role":"problem-context"}],"relatedBlocks":[{"id":"b-cc5c9bd76476d98ad9b677ec","url":"#b-cc5c9bd76476d98ad9b677ec","relation":"figure-and-explanation"}]},{"id":"b-830fe91eccf61d41fb31db19","type":"p","scope":"mask-branch-target","url":"https://codingai-lec03.pages.dev/course/blocks/mask-branch-target.html#b-830fe91eccf61d41fb31db19","content":"本课程并没有在 LAB11 训练完整 Mask R-CNN；LAB11 是三类语义分割。Mask R-CNN 的训练证据来自原论文，坐标、局部目标和插值机制则可由 LAB03 的小算例独立检验。","sha256":"5cea2ec3cb0e60da6a03638a2746272cc387521d0c8b748c2e2c06412a9e1640","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"instance-mask-identity","title":"怎样让每一份轮廓属于一个明确对象？"},"page":{"id":"reading-opening","title":"一个 RoI 的三项输出，为什么不是三个类别标签","number":1},"sectionTitle":"一个 RoI 的三项输出，为什么不是三个类别标签","currentQuestion":"一个 RoI 的三项输出，为什么不是三个类别标签","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ca2129f787710fafd4f9d866","nextBlock":"b-cc5c9bd76476d98ad9b677ec","prerequisites":[{"id":"b-eb5e44e634975dbec12ce39c","url":"#b-eb5e44e634975dbec12ce39c","role":"section-title"},{"id":"b-d445c7d76675b3f20a9c822b","url":"#b-d445c7d76675b3f20a9c822b","role":"problem-context"}],"relatedBlocks":[{"id":"b-cc5c9bd76476d98ad9b677ec","url":"#b-cc5c9bd76476d98ad9b677ec","relation":"figure-and-explanation"}]},{"id":"b-cc5c9bd76476d98ad9b677ec","type":"figure","scope":"mask-branch-target","url":"https://codingai-lec03.pages.dev/course/blocks/mask-branch-target.html#b-cc5c9bd76476d98ad9b677ec","content":"图内标签与数值：Lecture 03 课堂图解 P40；原图 + 当前 RoI；共享 F；RoI 读取；二维区域特征 F_RoI；类别：杯子；框：位置修正；原图局部：仅用于对应对象；实例；二值 mask；教学目标示意 · 非模型预测；三项任务读取共享 F 中的区域特征；照片仅作参考，mask 保持二维局部布局。\n\n共享 F 读取出的二维区域特征进入三项任务头；照片独立作参考，掩码是教学目标示意。 查看对应图解状态\n\n- 原图 + 当前 RoI\n- 共享 F\n- RoI 读取\n- 二维区域特征 F_RoI\n- 类别：杯子\n- 框：位置修正\n- 原图局部：仅用于对应对象\n- 实例\n- 二值 mask\n- 教学目标示意 · 非模型预测\n- 三项任务读取共享 F 中的区域特征；照片仅作参考，mask 保持二维局部布局。共享 F 读取出的二维区域特征进入三项任务头；照片独立作参考，掩码是教学目标示意。 [查看对应图解状态](/#/l3-r2-p40/1)[查看完整图解](/#/l3-r2-p40/1)","sha256":"33b3dd7a7fca0aa326b276734d6d324ea0ef699894580ba29b7682a78029df04","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"instance-mask-identity","title":"怎样让每一份轮廓属于一个明确对象？"},"page":{"id":"reading-opening","title":"一个 RoI 的三项输出，为什么不是三个类别标签","number":1},"sectionTitle":"一个 RoI 的三项输出，为什么不是三个类别标签","currentQuestion":"一个 RoI 的三项输出，为什么不是三个类别标签","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"共享 F 读取出的二维区域特征进入三项任务头；照片独立作参考，掩码是教学目标示意。 查看对应图解状态","labels":["原图 + 当前 RoI","共享 F","RoI 读取","二维区域特征 F_RoI","类别：杯子","框：位置修正","原图局部：仅用于对应对象","实例","二值 mask","教学目标示意 · 非模型预测","三项任务读取共享 F 中的区域特征；照片仅作参考，mask 保持二维局部布局。"],"description":"共享 F 读取出的二维区域特征进入三项任务头；照片独立作参考，掩码是教学目标示意。 查看对应图解状态原图 + 当前 RoI共享 FRoI 读取二维区域特征 F_RoI类别：杯子框：位置修正原图局部：仅用于对应对象实例二值 mask教学目标示意 · 非模型预测三项任务读取共享 F 中的区域特征；照片仅作参考，mask 保持二维局部布局。","evidenceIdentity":"共享 F 读取出的二维区域特征进入三项任务头；照片独立作参考，掩码是教学目标示意。 查看对应图解状态","images":[{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""},{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""}]},"previousBlock":"b-830fe91eccf61d41fb31db19","nextBlock":"b-0feadac72e948ffe131963c7","prerequisites":[{"id":"b-eb5e44e634975dbec12ce39c","url":"#b-eb5e44e634975dbec12ce39c","role":"section-title"},{"id":"b-d445c7d76675b3f20a9c822b","url":"#b-d445c7d76675b3f20a9c822b","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-0feadac72e948ffe131963c7","type":"h2","scope":"mask-branch-target","url":"https://codingai-lec03.pages.dev/course/blocks/mask-branch-target.html#b-0feadac72e948ffe131963c7","content":"## 继续阅读与实验","sha256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"instance-mask-identity","title":"怎样让每一份轮廓属于一个明确对象？"},"page":{"id":"reading-opening","title":"一个 RoI 的三项输出，为什么不是三个类别标签","number":1},"sectionTitle":"一个 RoI 的三项输出，为什么不是三个类别标签","currentQuestion":"一个 RoI 的三项输出，为什么不是三个类别标签","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-cc5c9bd76476d98ad9b677ec","nextBlock":null,"prerequisites":[{"id":"b-eb5e44e634975dbec12ce39c","url":"#b-eb5e44e634975dbec12ce39c","role":"section-title"},{"id":"b-d445c7d76675b3f20a9c822b","url":"#b-d445c7d76675b3f20a9c822b","role":"problem-context"}],"relatedBlocks":[{"id":"b-cc5c9bd76476d98ad9b677ec","url":"#b-cc5c9bd76476d98ad9b677ec","relation":"figure-and-explanation"}]}]}