{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/blocks/fast-rcnn-two-heads.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"共享特征怎样收到每个区域的训练反馈","url":"https://codingai-lec03.pages.dev/course/blocks/fast-rcnn-two-heads.html","markdown":"https://codingai-lec03.pages.dev/course/blocks/fast-rcnn-two-heads.md","pages":[{"id":"reading-opening","title":"共享特征怎样收到每个区域的训练反馈","unit":"类别与位置怎样成为可学习的区域目标？","unitId":"learnable-box-prediction","chapter":"II","number":1}],"aliases":[{"oldId":"b-ea1e91989c9285649ee7545b","targetId":"b-ea1e91989c9285649ee7545b","page":"reading-opening","operation":"move","beforeSHA256":"deb20c034fc271a6452aad283fd1175b5eeaac5a163afe41b583d29381035446","afterSHA256":"deb20c034fc271a6452aad283fd1175b5eeaac5a163afe41b583d29381035446","reason":"Same source body in authored semantic page"},{"oldId":"b-f12f729c1e1a406903dba922","targetId":"b-f12f729c1e1a406903dba922","page":"reading-opening","operation":"move","beforeSHA256":"a10e5e2e6d482cbef9ce4f17a89bba1ad5570d21cbc0682a29ab29fd083db01d","afterSHA256":"a10e5e2e6d482cbef9ce4f17a89bba1ad5570d21cbc0682a29ab29fd083db01d","reason":"Same source body in authored semantic page"},{"oldId":"b-947c881d7511c171958dfd15","targetId":"b-947c881d7511c171958dfd15","page":"reading-opening","operation":"move","beforeSHA256":"87e1ddc05e492c9328d3500ad1c8e4b58ae471d196f166627f04c93bbb95942a","afterSHA256":"87e1ddc05e492c9328d3500ad1c8e4b58ae471d196f166627f04c93bbb95942a","reason":"Same source body in authored semantic page"},{"oldId":"b-b7f28892f4554297bd7c87af","targetId":"b-b7f28892f4554297bd7c87af","page":"reading-opening","operation":"move","beforeSHA256":"5d15802c85bd3afc462f6a460ea201d77b7264fb0fc52bb2487e477882bb056c","afterSHA256":"5d15802c85bd3afc462f6a460ea201d77b7264fb0fc52bb2487e477882bb056c","reason":"Same source body in authored semantic page"},{"oldId":"b-c100b47f8a1e0620dd936978","targetId":"b-c100b47f8a1e0620dd936978","page":"reading-opening","operation":"move","beforeSHA256":"279e8b655d392285cb663a9a1b5f511c83a2f241603a5f221d31282a48b80e01","afterSHA256":"279e8b655d392285cb663a9a1b5f511c83a2f241603a5f221d31282a48b80e01","reason":"Same source body in authored semantic page"},{"oldId":"b-09111505b5e2b0bdabf2ca8c","targetId":"b-09111505b5e2b0bdabf2ca8c","page":"reading-opening","operation":"move","beforeSHA256":"fd190ed6180e2e409aac31bdda229fafdfea61aff28409486ceabf54248099b1","afterSHA256":"fd190ed6180e2e409aac31bdda229fafdfea61aff28409486ceabf54248099b1","reason":"Same source body in authored semantic page"},{"oldId":"b-7805720a1d06d87413aede13","targetId":"b-7805720a1d06d87413aede13","page":"reading-opening","operation":"rewrite","beforeSHA256":"8a0500b765083eed6d57a9b3f41492adc0e8aabf14dd0ea1a96a1c4cc0f81e3e","afterSHA256":"fa427a12e36a64ed376b11576daea2ce82af5219ce8f2579fbfadcbb84a97862","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-3b8d288160515f2c17c3dc13","targetId":"b-3b8d288160515f2c17c3dc13","page":"reading-opening","operation":"move","beforeSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","afterSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","reason":"Same source body in authored semantic page"}],"blocks":[{"id":"b-ea1e91989c9285649ee7545b","type":"h1","scope":"fast-rcnn-two-heads","url":"https://codingai-lec03.pages.dev/course/blocks/fast-rcnn-two-heads.html#b-ea1e91989c9285649ee7545b","content":"# 共享特征怎样收到每个区域的训练反馈","sha256":"deb20c034fc271a6452aad283fd1175b5eeaac5a163afe41b583d29381035446","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reading-opening","title":"共享特征怎样收到每个区域的训练反馈","number":1},"sectionTitle":"共享特征怎样收到每个区域的训练反馈","currentQuestion":"共享特征怎样收到每个区域的训练反馈","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-f12f729c1e1a406903dba922","prerequisites":[{"id":"b-f12f729c1e1a406903dba922","url":"#b-f12f729c1e1a406903dba922","role":"problem-context"}],"relatedBlocks":[{"id":"b-7805720a1d06d87413aede13","url":"#b-7805720a1d06d87413aede13","relation":"figure-and-explanation"}]},{"id":"b-f12f729c1e1a406903dba922","type":"p","scope":"fast-rcnn-two-heads","url":"https://codingai-lec03.pages.dev/course/blocks/fast-rcnn-two-heads.html#b-f12f729c1e1a406903dba922","content":"Fast R-CNN 的改动包括共享卷积、区域读取和类别/位置反馈共同训练。图上的多个区域仍各有答案，合并的是可共用计算。先指出损失从哪里回到同一骨干，再理解为什么外部候选成为下一项瓶颈。","sha256":"a10e5e2e6d482cbef9ce4f17a89bba1ad5570d21cbc0682a29ab29fd083db01d","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reading-opening","title":"共享特征怎样收到每个区域的训练反馈","number":1},"sectionTitle":"共享特征怎样收到每个区域的训练反馈","currentQuestion":"共享特征怎样收到每个区域的训练反馈","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ea1e91989c9285649ee7545b","nextBlock":"b-947c881d7511c171958dfd15","prerequisites":[{"id":"b-ea1e91989c9285649ee7545b","url":"#b-ea1e91989c9285649ee7545b","role":"section-title"}],"relatedBlocks":[{"id":"b-7805720a1d06d87413aede13","url":"#b-7805720a1d06d87413aede13","relation":"figure-and-explanation"}]},{"id":"b-947c881d7511c171958dfd15","type":"p","scope":"fast-rcnn-two-heads","url":"https://codingai-lec03.pages.dev/course/blocks/fast-rcnn-two-heads.html#b-947c881d7511c171958dfd15","content":"从整图特征取到每个 RoI 后，分类分支判断类别或背景，框分支预测相对位置修正。设某个区域属于杯子，损失一部分来自“杯子”概率是否足够大，另一部分来自中心、宽高偏移是否接近匹配目标。背景区域参与分类，通常不承担指向某个对象的位置回归。","sha256":"87e1ddc05e492c9328d3500ad1c8e4b58ae471d196f166627f04c93bbb95942a","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reading-opening","title":"共享特征怎样收到每个区域的训练反馈","number":1},"sectionTitle":"共享特征怎样收到每个区域的训练反馈","currentQuestion":"共享特征怎样收到每个区域的训练反馈","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f12f729c1e1a406903dba922","nextBlock":"b-b7f28892f4554297bd7c87af","prerequisites":[{"id":"b-ea1e91989c9285649ee7545b","url":"#b-ea1e91989c9285649ee7545b","role":"section-title"},{"id":"b-f12f729c1e1a406903dba922","url":"#b-f12f729c1e1a406903dba922","role":"problem-context"}],"relatedBlocks":[{"id":"b-7805720a1d06d87413aede13","url":"#b-7805720a1d06d87413aede13","relation":"figure-and-explanation"}]},{"id":"b-b7f28892f4554297bd7c87af","type":"p","scope":"fast-rcnn-two-heads","url":"https://codingai-lec03.pages.dev/course/blocks/fast-rcnn-two-heads.html#b-b7f28892f4554297bd7c87af","content":"把一批区域的损失相加后，梯度沿各自的区域读出回到共享骨干。共享的参数会收到来自不同区域、不同任务的反馈；它并不要求所有区域有相同答案。正如同一把尺子测不同物体可以得到不同长度，同一个预测器处理不同区域特征也会给出不同结果。","sha256":"5d15802c85bd3afc462f6a460ea201d77b7264fb0fc52bb2487e477882bb056c","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reading-opening","title":"共享特征怎样收到每个区域的训练反馈","number":1},"sectionTitle":"共享特征怎样收到每个区域的训练反馈","currentQuestion":"共享特征怎样收到每个区域的训练反馈","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-947c881d7511c171958dfd15","nextBlock":"b-c100b47f8a1e0620dd936978","prerequisites":[{"id":"b-ea1e91989c9285649ee7545b","url":"#b-ea1e91989c9285649ee7545b","role":"section-title"},{"id":"b-f12f729c1e1a406903dba922","url":"#b-f12f729c1e1a406903dba922","role":"problem-context"}],"relatedBlocks":[{"id":"b-7805720a1d06d87413aede13","url":"#b-7805720a1d06d87413aede13","relation":"figure-and-explanation"}]},{"id":"b-c100b47f8a1e0620dd936978","type":"p","scope":"fast-rcnn-two-heads","url":"https://codingai-lec03.pages.dev/course/blocks/fast-rcnn-two-heads.html#b-c100b47f8a1e0620dd936978","content":"Fast R-CNN 保留了外部候选来源，但把共享卷积、RoI 读取、分类和框回归组织进更紧凑的训练流程。卷积不再为每个重叠区域重复做，候选提取的时间便更加突出。下一步自然不是继续增加分类头，而是让候选也从这张 F 中学习。","sha256":"279e8b655d392285cb663a9a1b5f511c83a2f241603a5f221d31282a48b80e01","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reading-opening","title":"共享特征怎样收到每个区域的训练反馈","number":1},"sectionTitle":"共享特征怎样收到每个区域的训练反馈","currentQuestion":"共享特征怎样收到每个区域的训练反馈","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b7f28892f4554297bd7c87af","nextBlock":"b-09111505b5e2b0bdabf2ca8c","prerequisites":[{"id":"b-ea1e91989c9285649ee7545b","url":"#b-ea1e91989c9285649ee7545b","role":"section-title"},{"id":"b-f12f729c1e1a406903dba922","url":"#b-f12f729c1e1a406903dba922","role":"problem-context"}],"relatedBlocks":[{"id":"b-7805720a1d06d87413aede13","url":"#b-7805720a1d06d87413aede13","relation":"figure-and-explanation"}]},{"id":"b-09111505b5e2b0bdabf2ca8c","type":"p","scope":"fast-rcnn-two-heads","url":"https://codingai-lec03.pages.dev/course/blocks/fast-rcnn-two-heads.html#b-09111505b5e2b0bdabf2ca8c","content":"理解这一页时，能够在图上指出“两个区域各自的输出”和“同一个被共同更新的骨干”，就已经抓住它的机制。历史多阶段实现细节与训练采样放在 CLS07，而不加回主讲页。","sha256":"fd190ed6180e2e409aac31bdda229fafdfea61aff28409486ceabf54248099b1","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reading-opening","title":"共享特征怎样收到每个区域的训练反馈","number":1},"sectionTitle":"共享特征怎样收到每个区域的训练反馈","currentQuestion":"共享特征怎样收到每个区域的训练反馈","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c100b47f8a1e0620dd936978","nextBlock":"b-7805720a1d06d87413aede13","prerequisites":[{"id":"b-ea1e91989c9285649ee7545b","url":"#b-ea1e91989c9285649ee7545b","role":"section-title"},{"id":"b-f12f729c1e1a406903dba922","url":"#b-f12f729c1e1a406903dba922","role":"problem-context"}],"relatedBlocks":[{"id":"b-7805720a1d06d87413aede13","url":"#b-7805720a1d06d87413aede13","relation":"figure-and-explanation"}]},{"id":"b-7805720a1d06d87413aede13","type":"figure","scope":"fast-rcnn-two-heads","url":"https://codingai-lec03.pages.dev/course/blocks/fast-rcnn-two-heads.html#b-7805720a1d06d87413aede13","content":"图内标签与数值：Lecture 03 课堂图解 P31；共享卷积特征；RoI Pool；类别；框修正；类别；框修正；两项任务反馈到共享骨干\n\n多个区域的类别与回归损失共同更新共享骨干。 查看对应图解状态\n\n- 共享卷积特征\n- RoI Pool\n- 类别\n- 框修正\n- 类别\n- 框修正\n- 两项任务反馈到共享骨干多个区域的类别与回归损失共同更新共享骨干。 [查看对应图解状态](/#/l3-r2-p31/1)[查看完整图解](/#/l3-r2-p31/1)","sha256":"fa427a12e36a64ed376b11576daea2ce82af5219ce8f2579fbfadcbb84a97862","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reading-opening","title":"共享特征怎样收到每个区域的训练反馈","number":1},"sectionTitle":"共享特征怎样收到每个区域的训练反馈","currentQuestion":"共享特征怎样收到每个区域的训练反馈","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"多个区域的类别与回归损失共同更新共享骨干。 查看对应图解状态","labels":["共享卷积特征","RoI Pool","类别","框修正","类别","框修正","两项任务反馈到共享骨干"],"description":"多个区域的类别与回归损失共同更新共享骨干。 查看对应图解状态共享卷积特征RoI Pool类别框修正类别框修正两项任务反馈到共享骨干","evidenceIdentity":"多个区域的类别与回归损失共同更新共享骨干。 查看对应图解状态","images":[{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""}]},"previousBlock":"b-09111505b5e2b0bdabf2ca8c","nextBlock":"b-3b8d288160515f2c17c3dc13","prerequisites":[{"id":"b-ea1e91989c9285649ee7545b","url":"#b-ea1e91989c9285649ee7545b","role":"section-title"},{"id":"b-f12f729c1e1a406903dba922","url":"#b-f12f729c1e1a406903dba922","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-3b8d288160515f2c17c3dc13","type":"h2","scope":"fast-rcnn-two-heads","url":"https://codingai-lec03.pages.dev/course/blocks/fast-rcnn-two-heads.html#b-3b8d288160515f2c17c3dc13","content":"## 继续阅读与实验","sha256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"learnable-box-prediction","title":"类别与位置怎样成为可学习的区域目标？"},"page":{"id":"reading-opening","title":"共享特征怎样收到每个区域的训练反馈","number":1},"sectionTitle":"共享特征怎样收到每个区域的训练反馈","currentQuestion":"共享特征怎样收到每个区域的训练反馈","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7805720a1d06d87413aede13","nextBlock":null,"prerequisites":[{"id":"b-ea1e91989c9285649ee7545b","url":"#b-ea1e91989c9285649ee7545b","role":"section-title"},{"id":"b-f12f729c1e1a406903dba922","url":"#b-f12f729c1e1a406903dba922","role":"problem-context"}],"relatedBlocks":[{"id":"b-7805720a1d06d87413aede13","url":"#b-7805720a1d06d87413aede13","relation":"figure-and-explanation"}]}]}